{"text": " You have seen the thumbnails by now. KimiK3 Plus Claude Code, free, local, open source. Point the best coding harness on the planet at an open model and stop paying Anthropic forever. It sounds too good, and the honest answer is that most of it is real, and one word in that title is quietly doing a lot of lifting, because here is what is true. KimiK3 is genuinely open. The weights are yours. It is a fraction of the price of the frontier, and it does slot into Claude Code, the exact agent loop most people already trust, but the word local is where the fantasy and the hardware collide, and I want to show you exactly where that line sits before you rip out your setup, so here is the plan. What KimiK3 actually is, how you wire it into Claude Code in a few minutes, and the honest truth about free versus local versus what your machine can really run. Start with what shipped. KimiK3 is an open-weight model from Moonshot AI, released this month under a modified MIT license. It is a mixture of experts model, 2.8 trillion parameters total, but only 16 of its 896 experts fire for any given token, so it is enormous on disk and lean when it actually runs. It carries a 1 million token context window, it reads images and video natively, and its launch materials put it head-to-head with Claude Fable 5 on coding and agent benchmarks. That is the headline. Here is what the headline leaves out. On raw intelligence, K3 does not beat Fable 5. Independent testing from artificial analysis puts it around 3 points behind the leader on the overall intelligence index, but 3 points behind the best model on earth, while being open and roughly a third of the price, is a sentence that did not exist a month ago. And on the axis that matters for a coding agent, it is right in the mix. It beats Opus 4.8 and GPT 5.5 on several coding benchmarks, and wins 2 of 6 outright in Moonshot's own evaluation. On GDP Vol, a benchmark of real economically valuable tasks, it posts an ELO above Opus 4.8. So the gap is real on the hardest novel problems, and basically invisible on the everyday 90% of the job. Now, why does this pair with Claude Code specifically? Because Claude Code is just a harness. Strip it down, and it is a loop. Gather context, take an action through a tool call, verify the result, repeat until the task is done. The file edits, the shell commands, the plan and verify cycle, all of it rides on that loop. It was tuned for Claude, but the loop does not actually care which model sits underneath, as long as that model is good at two things, calling tools in the right format, and following instructions without drifting. And K3 was built for exactly those two things. Moonshot trained it for long horizon agent work, the kind that runs for an hour instead of answering in one shot. Read the repo, make a plan, edit 20 files, run the tests, read the failures, fix what broke, and try again. That is the loop where most models drift and fall apart around step 10. It is also the exact loop Claude Code runs, and the exact loop K3 was trained to survive. There is even a K3 variant, K3 Swarm, built to fan out many parallel copies on one big job, the same shape as a coding agent spawning a dozen subtasks at once. So how do you actually wire it up? This is the part that is genuinely a few minutes. Claude Code talks to an Anthropic-style endpoint, and it reads that endpoint from environment variables. So you point those variables at a provider serving Kimi K3 instead of at Anthropic. Set the base URL, set the API key, name the K3 model, and launch Claude Code as you normally would. The harness never knows the difference. It sends the same tool definitions, the same system prompt, the same growing context, and K3 answers in the same shape Claude would. Nothing else in your workflow changes. If you use a router like OpenRouter, it is even shorter. K3 is already listed, priced exactly as Moonshot published, so you set one base URL, drop in your router key, and you are running in a single config change. And this is where free earns a real, if partial, meaning. Some providers in the Kimi platform itself run promotional free tiers and trial credits at launch. So for light use, you genuinely can drive Claude Code with K3 without paying for a while. Not free forever, and not your own hardware, but the entry cost really can be zero to start. That is the honest version of the free in the title. One practical note before you switch everything over. Test it on a real task first, not a toy prompt. Point K3 at an actual bug in an actual repo, and watch it run a full loop. What you are checking is not whether it can write a function. It is whether it holds the thread across 10 tool calls without losing the plan, and that only shows up on real work. And here is what that unlocks. The token math flips hard. Kimi K3 runs at $3 per million input and $15 per million output. Claude Fable 5 charges $10 and $50. So you are paying roughly a third for a model that lands three points behind. And if your workflow reuses context the way every coding agent does, rereading the same files over and over, cached input drops to $0.30 a million, a tenth of the sticker price. Put it in real terms. A heavy day of agentic coding can burn 20 million tokens on Fable 5's output rate, a bill that makes you flinch. On K3, with caching on the input, the same work is closer to lunch money. Now, the word in the title I promise to be honest about. Local. Because free and open source are true, but local is where the hardware laughs at you. Kimi K3 is 2.8 trillion parameters. At the 4-bit format Moonshot ships, that is about 1.4 terabytes just for the raw weights. The largest Mac studio tops out at 512 gigabytes. 4H100 give you 320. Even a single 8GPU B200 server, at 1.44 terabytes, barely clears the floor with nothing left for context or runtime. Running the full K3 on your own machine is not a desktop project. It is a data center project. So when a thumbnail says local, what it almost always means is running it through a provider on someone else's GPUs. The model is open. Your laptop still cannot hold it. Let me be fair, though, because there is a real local story here, just a smaller one. Open waits mean you are not locked to a vendor. A company with the hardware, or a rented server genuinely can self-host K3 for privacy or compliance, and no policy change can switch it off. Remember, Fable 5 got pulled offline for days over an export control scare. You cannot switch off a file already sitting on 10,000 hard drives. That durability is real, and it is the actual meaning of open here, even if it never touches your desk. And the honest downside does not stop at hardware. K3 trails the frontier on the hardest, most novel problems, and you will feel that gap on genuinely tricky architecture work. Its safety and refusal behavior is different from Claude's, tuned by a different lab with different defaults. And the polish of the Claude plus Claude code experience, the little ways the harness and the model were co-designed, is not something a swapped-in model perfectly reproduces. It is close, not identical. To be fair to just paying for Claude, that co-design is worth something. If your work is the hardest 10%, novel systems, subtle multi-file reasoning, the frontier model earns its price, and the smoothest experience is still Claude code pointed at Claude. Nobody optimized that pairing harder than Anthropic. So here is the honest verdict. If you run a lot of agentic coding, refactors, tests, glue code, the everyday 90%, Kimi K3 in Claude code is close to a cheat code. You keep the harness you trust, and cut your bill to a third or less. Use K3 for volume. Keep a frontier model one config flag away for the problems that actually need it. And treat local as run it through a provider, unless you literally own a GPU server, because the full weights will never fit on a laptop. Four numbers to walk away with. 2.8 trillion parameters. Three points behind Fable 5 on intelligence. Three dollars in, 15 out, a third of the frontier price, dropping to 30 cents cash. And 1.4 terabytes of weights, the number that decides whether local means your machine or someone else's. If this turned a hype thumbnail into an actual setup, you can run and a bill you can cut. Subscribe. This channel takes apart every open model and every local AI claim exactly like this. Receipts first, verdict second. I will see you in the next one.", "segments": [{"id": 0, "seek": 0, "start": 0.0, "end": 6.5, "text": " You have seen the thumbnails by now. KimiK3 Plus Claude Code, free, local, open source.", "tokens": [50365, 509, 362, 1612, 264, 46987, 538, 586, 13, 5652, 72, 42, 18, 7721, 12947, 2303, 15549, 11, 1737, 11, 2654, 11, 1269, 4009, 13, 50690], "temperature": 0, "avg_logprob": -0.1499483157426883, "compression_ratio": 1.6036363636363635, "no_speech_prob": 1.107061599510406e-12}, {"id": 1, "seek": 0, "start": 7.0200000000000005, "end": 11.540000000000001, "text": " Point the best coding harness on the planet at an open model and stop paying Anthropic forever.", "tokens": [50716, 12387, 264, 1151, 17720, 19700, 322, 264, 5054, 412, 364, 1269, 2316, 293, 1590, 6229, 12727, 39173, 5680, 13, 50942], "temperature": 0, "avg_logprob": -0.1499483157426883, "compression_ratio": 1.6036363636363635, "no_speech_prob": 1.107061599510406e-12}, {"id": 2, "seek": 0, "start": 12.06, "end": 17.72, "text": " It sounds too good, and the honest answer is that most of it is real, and one word in that title is", "tokens": [50968, 467, 3263, 886, 665, 11, 293, 264, 3245, 1867, 307, 300, 881, 295, 309, 307, 957, 11, 293, 472, 1349, 294, 300, 4876, 307, 51251], "temperature": 0, "avg_logprob": -0.1499483157426883, "compression_ratio": 1.6036363636363635, "no_speech_prob": 1.107061599510406e-12}, {"id": 3, "seek": 0, "start": 17.72, "end": 23.84, "text": " quietly doing a lot of lifting, because here is what is true. KimiK3 is genuinely open.", "tokens": [51251, 19141, 884, 257, 688, 295, 15798, 11, 570, 510, 307, 437, 307, 2074, 13, 5652, 72, 42, 18, 307, 17839, 1269, 13, 51557], "temperature": 0, "avg_logprob": -0.1499483157426883, "compression_ratio": 1.6036363636363635, "no_speech_prob": 1.107061599510406e-12}, {"id": 4, "seek": 0, "start": 24.28, "end": 27.82, "text": " The weights are yours. It is a fraction of the price of the frontier,", "tokens": [51579, 440, 17443, 366, 6342, 13, 467, 307, 257, 14135, 295, 264, 3218, 295, 264, 35853, 11, 51756], "temperature": 0, "avg_logprob": -0.1499483157426883, "compression_ratio": 1.6036363636363635, "no_speech_prob": 1.107061599510406e-12}, {"id": 5, "seek": 2782, "start": 27.82, "end": 33.06, "text": " and it does slot into Claude Code, the exact agent loop most people already trust,", "tokens": [50365, 293, 309, 775, 14747, 666, 12947, 2303, 15549, 11, 264, 1900, 9461, 6367, 881, 561, 1217, 3361, 11, 50627], "temperature": 0, "avg_logprob": -0.13572913642943374, "compression_ratio": 1.6769759450171822, "no_speech_prob": 6.983293609173769e-13}, {"id": 6, "seek": 2782, "start": 33.28, "end": 36.42, "text": " but the word local is where the fantasy and the hardware collide,", "tokens": [50638, 457, 264, 1349, 2654, 307, 689, 264, 13861, 293, 264, 8837, 49093, 11, 50795], "temperature": 0, "avg_logprob": -0.13572913642943374, "compression_ratio": 1.6769759450171822, "no_speech_prob": 6.983293609173769e-13}, {"id": 7, "seek": 2782, "start": 36.72, "end": 41.76, "text": " and I want to show you exactly where that line sits before you rip out your setup, so here is the plan.", "tokens": [50810, 293, 286, 528, 281, 855, 291, 2293, 689, 300, 1622, 12696, 949, 291, 12782, 484, 428, 8657, 11, 370, 510, 307, 264, 1393, 13, 51062], "temperature": 0, "avg_logprob": -0.13572913642943374, "compression_ratio": 1.6769759450171822, "no_speech_prob": 6.983293609173769e-13}, {"id": 8, "seek": 2782, "start": 41.980000000000004, "end": 46.379999999999995, "text": " What KimiK3 actually is, how you wire it into Claude Code in a few minutes,", "tokens": [51073, 708, 5652, 72, 42, 18, 767, 307, 11, 577, 291, 6234, 309, 666, 12947, 2303, 15549, 294, 257, 1326, 2077, 11, 51293], "temperature": 0, "avg_logprob": -0.13572913642943374, "compression_ratio": 1.6769759450171822, "no_speech_prob": 6.983293609173769e-13}, {"id": 9, "seek": 2782, "start": 46.760000000000005, "end": 50.82, "text": " and the honest truth about free versus local versus what your machine can really run.", "tokens": [51312, 293, 264, 3245, 3494, 466, 1737, 5717, 2654, 5717, 437, 428, 3479, 393, 534, 1190, 13, 51515], "temperature": 0, "avg_logprob": -0.13572913642943374, "compression_ratio": 1.6769759450171822, "no_speech_prob": 6.983293609173769e-13}, {"id": 10, "seek": 2782, "start": 51.2, "end": 55.3, "text": " Start with what shipped. KimiK3 is an open-weight model from Moonshot AI,", "tokens": [51534, 6481, 365, 437, 25312, 13, 5652, 72, 42, 18, 307, 364, 1269, 12, 12329, 2316, 490, 3335, 892, 12194, 7318, 11, 51739], "temperature": 0, "avg_logprob": -0.13572913642943374, "compression_ratio": 1.6769759450171822, "no_speech_prob": 6.983293609173769e-13}, {"id": 11, "seek": 5530, "start": 55.3, "end": 58.22, "text": " released this month under a modified MIT license.", "tokens": [50365, 4736, 341, 1618, 833, 257, 15873, 13100, 10476, 13, 50511], "temperature": 0, "avg_logprob": -0.17563652992248535, "compression_ratio": 1.5321428571428573, "no_speech_prob": 1.5132538234638448e-12}, {"id": 12, "seek": 5530, "start": 58.54, "end": 62.9, "text": " It is a mixture of experts model, 2.8 trillion parameters total,", "tokens": [50527, 467, 307, 257, 9925, 295, 8572, 2316, 11, 568, 13, 23, 18723, 9834, 3217, 11, 50745], "temperature": 0, "avg_logprob": -0.17563652992248535, "compression_ratio": 1.5321428571428573, "no_speech_prob": 1.5132538234638448e-12}, {"id": 13, "seek": 5530, "start": 63.31999999999999, "end": 67.56, "text": " but only 16 of its 896 experts fire for any given token,", "tokens": [50766, 457, 787, 3165, 295, 1080, 1649, 22962, 8572, 2610, 337, 604, 2212, 14862, 11, 50978], "temperature": 0, "avg_logprob": -0.17563652992248535, "compression_ratio": 1.5321428571428573, "no_speech_prob": 1.5132538234638448e-12}, {"id": 14, "seek": 5530, "start": 67.94, "end": 71.02, "text": " so it is enormous on disk and lean when it actually runs.", "tokens": [50997, 370, 309, 307, 11322, 322, 12355, 293, 11659, 562, 309, 767, 6676, 13, 51151], "temperature": 0, "avg_logprob": -0.17563652992248535, "compression_ratio": 1.5321428571428573, "no_speech_prob": 1.5132538234638448e-12}, {"id": 15, "seek": 5530, "start": 71.42, "end": 75.86, "text": " It carries a 1 million token context window, it reads images and video natively,", "tokens": [51171, 467, 16402, 257, 502, 2459, 14862, 4319, 4910, 11, 309, 15700, 5267, 293, 960, 8470, 356, 11, 51393], "temperature": 0, "avg_logprob": -0.17563652992248535, "compression_ratio": 1.5321428571428573, "no_speech_prob": 1.5132538234638448e-12}, {"id": 16, "seek": 5530, "start": 76.1, "end": 80.88, "text": " and its launch materials put it head-to-head with Claude Fable 5 on coding and agent benchmarks.", "tokens": [51405, 293, 1080, 4025, 5319, 829, 309, 1378, 12, 1353, 12, 1934, 365, 12947, 2303, 479, 712, 1025, 322, 17720, 293, 9461, 43751, 13, 51644], "temperature": 0, "avg_logprob": -0.17563652992248535, "compression_ratio": 1.5321428571428573, "no_speech_prob": 1.5132538234638448e-12}, {"id": 17, "seek": 5530, "start": 81.17999999999999, "end": 82.14, "text": " That is the headline.", "tokens": [51659, 663, 307, 264, 28380, 13, 51707], "temperature": 0, "avg_logprob": -0.17563652992248535, "compression_ratio": 1.5321428571428573, "no_speech_prob": 1.5132538234638448e-12}, {"id": 18, "seek": 8214, "start": 82.14, "end": 84.04, "text": " Here is what the headline leaves out.", "tokens": [50365, 1692, 307, 437, 264, 28380, 5510, 484, 13, 50460], "temperature": 0, "avg_logprob": -0.13723000040594138, "compression_ratio": 1.6317829457364341, "no_speech_prob": 2.0927185765545753e-12}, {"id": 19, "seek": 8214, "start": 84.48, "end": 87.94, "text": " On raw intelligence, K3 does not beat Fable 5.", "tokens": [50482, 1282, 8936, 7599, 11, 591, 18, 775, 406, 4224, 479, 712, 1025, 13, 50655], "temperature": 0, "avg_logprob": -0.13723000040594138, "compression_ratio": 1.6317829457364341, "no_speech_prob": 2.0927185765545753e-12}, {"id": 20, "seek": 8214, "start": 88.4, "end": 94.04, "text": " Independent testing from artificial analysis puts it around 3 points behind the leader on the overall intelligence index,", "tokens": [50678, 40310, 4997, 490, 11677, 5215, 8137, 309, 926, 805, 2793, 2261, 264, 5263, 322, 264, 4787, 7599, 8186, 11, 50960], "temperature": 0, "avg_logprob": -0.13723000040594138, "compression_ratio": 1.6317829457364341, "no_speech_prob": 2.0927185765545753e-12}, {"id": 21, "seek": 8214, "start": 94.26, "end": 96.68, "text": " but 3 points behind the best model on earth,", "tokens": [50971, 457, 805, 2793, 2261, 264, 1151, 2316, 322, 4120, 11, 51092], "temperature": 0, "avg_logprob": -0.13723000040594138, "compression_ratio": 1.6317829457364341, "no_speech_prob": 2.0927185765545753e-12}, {"id": 22, "seek": 8214, "start": 96.96000000000001, "end": 99.6, "text": " while being open and roughly a third of the price,", "tokens": [51106, 1339, 885, 1269, 293, 9810, 257, 2636, 295, 264, 3218, 11, 51238], "temperature": 0, "avg_logprob": -0.13723000040594138, "compression_ratio": 1.6317829457364341, "no_speech_prob": 2.0927185765545753e-12}, {"id": 23, "seek": 8214, "start": 99.88, "end": 102.26, "text": " is a sentence that did not exist a month ago.", "tokens": [51252, 307, 257, 8174, 300, 630, 406, 2514, 257, 1618, 2057, 13, 51371], "temperature": 0, "avg_logprob": -0.13723000040594138, "compression_ratio": 1.6317829457364341, "no_speech_prob": 2.0927185765545753e-12}, {"id": 24, "seek": 8214, "start": 102.46000000000001, "end": 106.18, "text": " And on the axis that matters for a coding agent, it is right in the mix.", "tokens": [51381, 400, 322, 264, 10298, 300, 7001, 337, 257, 17720, 9461, 11, 309, 307, 558, 294, 264, 2890, 13, 51567], "temperature": 0, "avg_logprob": -0.13723000040594138, "compression_ratio": 1.6317829457364341, "no_speech_prob": 2.0927185765545753e-12}, {"id": 25, "seek": 10618, "start": 106.18, "end": 111.34, "text": " It beats Opus 4.8 and GPT 5.5 on several coding benchmarks,", "tokens": [50365, 467, 16447, 12011, 301, 1017, 13, 23, 293, 26039, 51, 1025, 13, 20, 322, 2940, 17720, 43751, 11, 50623], "temperature": 0, "avg_logprob": -0.1478117648686204, "compression_ratio": 1.4603174603174602, "no_speech_prob": 2.288626536547933e-12}, {"id": 26, "seek": 10618, "start": 111.72000000000001, "end": 114.92, "text": " and wins 2 of 6 outright in Moonshot's own evaluation.", "tokens": [50642, 293, 10641, 568, 295, 1386, 35189, 294, 3335, 892, 12194, 311, 1065, 13344, 13, 50802], "temperature": 0, "avg_logprob": -0.1478117648686204, "compression_ratio": 1.4603174603174602, "no_speech_prob": 2.288626536547933e-12}, {"id": 27, "seek": 10618, "start": 115.46000000000001, "end": 119.22000000000001, "text": " On GDP Vol, a benchmark of real economically valuable tasks,", "tokens": [50829, 1282, 19599, 8911, 11, 257, 18927, 295, 957, 26811, 8263, 9608, 11, 51017], "temperature": 0, "avg_logprob": -0.1478117648686204, "compression_ratio": 1.4603174603174602, "no_speech_prob": 2.288626536547933e-12}, {"id": 28, "seek": 10618, "start": 119.52000000000001, "end": 122.0, "text": " it posts an ELO above Opus 4.8.", "tokens": [51032, 309, 12300, 364, 14426, 46, 3673, 12011, 301, 1017, 13, 23, 13, 51156], "temperature": 0, "avg_logprob": -0.1478117648686204, "compression_ratio": 1.4603174603174602, "no_speech_prob": 2.288626536547933e-12}, {"id": 29, "seek": 10618, "start": 122.28, "end": 124.78, "text": " So the gap is real on the hardest novel problems,", "tokens": [51170, 407, 264, 7417, 307, 957, 322, 264, 13158, 7613, 2740, 11, 51295], "temperature": 0, "avg_logprob": -0.1478117648686204, "compression_ratio": 1.4603174603174602, "no_speech_prob": 2.288626536547933e-12}, {"id": 30, "seek": 10618, "start": 125.02000000000001, "end": 128.24, "text": " and basically invisible on the everyday 90% of the job.", "tokens": [51307, 293, 1936, 14603, 322, 264, 7429, 4289, 4, 295, 264, 1691, 13, 51468], "temperature": 0, "avg_logprob": -0.1478117648686204, "compression_ratio": 1.4603174603174602, "no_speech_prob": 2.288626536547933e-12}, {"id": 31, "seek": 10618, "start": 128.44, "end": 131.06, "text": " Now, why does this pair with Claude Code specifically?", "tokens": [51478, 823, 11, 983, 775, 341, 6119, 365, 12947, 2303, 15549, 4682, 30, 51609], "temperature": 0, "avg_logprob": -0.1478117648686204, "compression_ratio": 1.4603174603174602, "no_speech_prob": 2.288626536547933e-12}, {"id": 32, "seek": 13106, "start": 131.06, "end": 133.34, "text": " Because Claude Code is just a harness.", "tokens": [50365, 1436, 12947, 2303, 15549, 307, 445, 257, 19700, 13, 50479], "temperature": 0, "avg_logprob": -0.17272296934637404, "compression_ratio": 1.6701388888888888, "no_speech_prob": 1.5735457641746775e-12}, {"id": 33, "seek": 13106, "start": 133.74, "end": 135.26, "text": " Strip it down, and it is a loop.", "tokens": [50499, 745, 8400, 309, 760, 11, 293, 309, 307, 257, 6367, 13, 50575], "temperature": 0, "avg_logprob": -0.17272296934637404, "compression_ratio": 1.6701388888888888, "no_speech_prob": 1.5735457641746775e-12}, {"id": 34, "seek": 13106, "start": 135.58, "end": 138.24, "text": " Gather context, take an action through a tool call,", "tokens": [50591, 39841, 4319, 11, 747, 364, 3069, 807, 257, 2290, 818, 11, 50724], "temperature": 0, "avg_logprob": -0.17272296934637404, "compression_ratio": 1.6701388888888888, "no_speech_prob": 1.5735457641746775e-12}, {"id": 35, "seek": 13106, "start": 138.7, "end": 141.46, "text": " verify the result, repeat until the task is done.", "tokens": [50747, 16888, 264, 1874, 11, 7149, 1826, 264, 5633, 307, 1096, 13, 50885], "temperature": 0, "avg_logprob": -0.17272296934637404, "compression_ratio": 1.6701388888888888, "no_speech_prob": 1.5735457641746775e-12}, {"id": 36, "seek": 13106, "start": 141.88, "end": 145.64000000000001, "text": " The file edits, the shell commands, the plan and verify cycle,", "tokens": [50906, 440, 3991, 41752, 11, 264, 8720, 16901, 11, 264, 1393, 293, 16888, 6586, 11, 51094], "temperature": 0, "avg_logprob": -0.17272296934637404, "compression_ratio": 1.6701388888888888, "no_speech_prob": 1.5735457641746775e-12}, {"id": 37, "seek": 13106, "start": 146.0, "end": 147.32, "text": " all of it rides on that loop.", "tokens": [51112, 439, 295, 309, 20773, 322, 300, 6367, 13, 51178], "temperature": 0, "avg_logprob": -0.17272296934637404, "compression_ratio": 1.6701388888888888, "no_speech_prob": 1.5735457641746775e-12}, {"id": 38, "seek": 13106, "start": 147.6, "end": 148.7, "text": " It was tuned for Claude,", "tokens": [51192, 467, 390, 10870, 337, 12947, 2303, 11, 51247], "temperature": 0, "avg_logprob": -0.17272296934637404, "compression_ratio": 1.6701388888888888, "no_speech_prob": 1.5735457641746775e-12}, {"id": 39, "seek": 13106, "start": 149.02, "end": 151.62, "text": " but the loop does not actually care which model sits underneath,", "tokens": [51263, 457, 264, 6367, 775, 406, 767, 1127, 597, 2316, 12696, 7223, 11, 51393], "temperature": 0, "avg_logprob": -0.17272296934637404, "compression_ratio": 1.6701388888888888, "no_speech_prob": 1.5735457641746775e-12}, {"id": 40, "seek": 13106, "start": 152.04, "end": 153.9, "text": " as long as that model is good at two things,", "tokens": [51414, 382, 938, 382, 300, 2316, 307, 665, 412, 732, 721, 11, 51507], "temperature": 0, "avg_logprob": -0.17272296934637404, "compression_ratio": 1.6701388888888888, "no_speech_prob": 1.5735457641746775e-12}, {"id": 41, "seek": 13106, "start": 154.34, "end": 155.82, "text": " calling tools in the right format,", "tokens": [51529, 5141, 3873, 294, 264, 558, 7877, 11, 51603], "temperature": 0, "avg_logprob": -0.17272296934637404, "compression_ratio": 1.6701388888888888, "no_speech_prob": 1.5735457641746775e-12}, {"id": 42, "seek": 13106, "start": 156.24, "end": 157.96, "text": " and following instructions without drifting.", "tokens": [51624, 293, 3480, 9415, 1553, 37973, 13, 51710], "temperature": 0, "avg_logprob": -0.17272296934637404, "compression_ratio": 1.6701388888888888, "no_speech_prob": 1.5735457641746775e-12}, {"id": 43, "seek": 15796, "start": 157.96, "end": 161.12, "text": " And K3 was built for exactly those two things.", "tokens": [50365, 400, 591, 18, 390, 3094, 337, 2293, 729, 732, 721, 13, 50523], "temperature": 0, "avg_logprob": -0.14103448095400473, "compression_ratio": 1.6226415094339623, "no_speech_prob": 1.6426270066371984e-12}, {"id": 44, "seek": 15796, "start": 161.48000000000002, "end": 163.8, "text": " Moonshot trained it for long horizon agent work,", "tokens": [50541, 3335, 892, 12194, 8895, 309, 337, 938, 18046, 9461, 589, 11, 50657], "temperature": 0, "avg_logprob": -0.14103448095400473, "compression_ratio": 1.6226415094339623, "no_speech_prob": 1.6426270066371984e-12}, {"id": 45, "seek": 15796, "start": 164.14000000000001, "end": 167.0, "text": " the kind that runs for an hour instead of answering in one shot.", "tokens": [50674, 264, 733, 300, 6676, 337, 364, 1773, 2602, 295, 13430, 294, 472, 3347, 13, 50817], "temperature": 0, "avg_logprob": -0.14103448095400473, "compression_ratio": 1.6226415094339623, "no_speech_prob": 1.6426270066371984e-12}, {"id": 46, "seek": 15796, "start": 167.38, "end": 170.84, "text": " Read the repo, make a plan, edit 20 files,", "tokens": [50836, 17604, 264, 49040, 11, 652, 257, 1393, 11, 8129, 945, 7098, 11, 51009], "temperature": 0, "avg_logprob": -0.14103448095400473, "compression_ratio": 1.6226415094339623, "no_speech_prob": 1.6426270066371984e-12}, {"id": 47, "seek": 15796, "start": 171.38, "end": 173.32, "text": " run the tests, read the failures,", "tokens": [51036, 1190, 264, 6921, 11, 1401, 264, 20774, 11, 51133], "temperature": 0, "avg_logprob": -0.14103448095400473, "compression_ratio": 1.6226415094339623, "no_speech_prob": 1.6426270066371984e-12}, {"id": 48, "seek": 15796, "start": 173.9, "end": 175.76000000000002, "text": " fix what broke, and try again.", "tokens": [51162, 3191, 437, 6902, 11, 293, 853, 797, 13, 51255], "temperature": 0, "avg_logprob": -0.14103448095400473, "compression_ratio": 1.6226415094339623, "no_speech_prob": 1.6426270066371984e-12}, {"id": 49, "seek": 15796, "start": 176.12, "end": 179.78, "text": " That is the loop where most models drift and fall apart around step 10.", "tokens": [51273, 663, 307, 264, 6367, 689, 881, 5245, 19699, 293, 2100, 4936, 926, 1823, 1266, 13, 51456], "temperature": 0, "avg_logprob": -0.14103448095400473, "compression_ratio": 1.6226415094339623, "no_speech_prob": 1.6426270066371984e-12}, {"id": 50, "seek": 15796, "start": 180.10000000000002, "end": 182.56, "text": " It is also the exact loop Claude Code runs,", "tokens": [51472, 467, 307, 611, 264, 1900, 6367, 12947, 2303, 15549, 6676, 11, 51595], "temperature": 0, "avg_logprob": -0.14103448095400473, "compression_ratio": 1.6226415094339623, "no_speech_prob": 1.6426270066371984e-12}, {"id": 51, "seek": 15796, "start": 182.84, "end": 185.48000000000002, "text": " and the exact loop K3 was trained to survive.", "tokens": [51609, 293, 264, 1900, 6367, 591, 18, 390, 8895, 281, 7867, 13, 51741], "temperature": 0, "avg_logprob": -0.14103448095400473, "compression_ratio": 1.6226415094339623, "no_speech_prob": 1.6426270066371984e-12}, {"id": 52, "seek": 18548, "start": 185.48, "end": 188.82, "text": " There is even a K3 variant, K3 Swarm,", "tokens": [50365, 821, 307, 754, 257, 591, 18, 17501, 11, 591, 18, 3926, 4452, 11, 50532], "temperature": 0, "avg_logprob": -0.16940590540568035, "compression_ratio": 1.586080586080586, "no_speech_prob": 1.920417600984048e-12}, {"id": 53, "seek": 18548, "start": 189.2, "end": 192.23999999999998, "text": " built to fan out many parallel copies on one big job,", "tokens": [50551, 3094, 281, 3429, 484, 867, 8952, 14341, 322, 472, 955, 1691, 11, 50703], "temperature": 0, "avg_logprob": -0.16940590540568035, "compression_ratio": 1.586080586080586, "no_speech_prob": 1.920417600984048e-12}, {"id": 54, "seek": 18548, "start": 192.56, "end": 195.98, "text": " the same shape as a coding agent spawning a dozen subtasks at once.", "tokens": [50719, 264, 912, 3909, 382, 257, 17720, 9461, 637, 35880, 257, 16654, 7257, 296, 1694, 412, 1564, 13, 50890], "temperature": 0, "avg_logprob": -0.16940590540568035, "compression_ratio": 1.586080586080586, "no_speech_prob": 1.920417600984048e-12}, {"id": 55, "seek": 18548, "start": 196.44, "end": 198.14, "text": " So how do you actually wire it up?", "tokens": [50913, 407, 577, 360, 291, 767, 6234, 309, 493, 30, 50998], "temperature": 0, "avg_logprob": -0.16940590540568035, "compression_ratio": 1.586080586080586, "no_speech_prob": 1.920417600984048e-12}, {"id": 56, "seek": 18548, "start": 198.45999999999998, "end": 200.64, "text": " This is the part that is genuinely a few minutes.", "tokens": [51014, 639, 307, 264, 644, 300, 307, 17839, 257, 1326, 2077, 13, 51123], "temperature": 0, "avg_logprob": -0.16940590540568035, "compression_ratio": 1.586080586080586, "no_speech_prob": 1.920417600984048e-12}, {"id": 57, "seek": 18548, "start": 201.2, "end": 203.54, "text": " Claude Code talks to an Anthropic-style endpoint,", "tokens": [51151, 12947, 2303, 15549, 6686, 281, 364, 12727, 39173, 12, 15014, 35795, 11, 51268], "temperature": 0, "avg_logprob": -0.16940590540568035, "compression_ratio": 1.586080586080586, "no_speech_prob": 1.920417600984048e-12}, {"id": 58, "seek": 18548, "start": 204.0, "end": 206.38, "text": " and it reads that endpoint from environment variables.", "tokens": [51291, 293, 309, 15700, 300, 35795, 490, 2823, 9102, 13, 51410], "temperature": 0, "avg_logprob": -0.16940590540568035, "compression_ratio": 1.586080586080586, "no_speech_prob": 1.920417600984048e-12}, {"id": 59, "seek": 18548, "start": 206.79999999999998, "end": 211.17999999999998, "text": " So you point those variables at a provider serving Kimi K3 instead of at Anthropic.", "tokens": [51431, 407, 291, 935, 729, 9102, 412, 257, 12398, 8148, 5652, 72, 591, 18, 2602, 295, 412, 12727, 39173, 13, 51650], "temperature": 0, "avg_logprob": -0.16940590540568035, "compression_ratio": 1.586080586080586, "no_speech_prob": 1.920417600984048e-12}, {"id": 60, "seek": 21118, "start": 211.18, "end": 216.04000000000002, "text": " Set the base URL, set the API key, name the K3 model,", "tokens": [50365, 8928, 264, 3096, 12905, 11, 992, 264, 9362, 2141, 11, 1315, 264, 591, 18, 2316, 11, 50608], "temperature": 0, "avg_logprob": -0.15942821732486587, "compression_ratio": 1.5170731707317073, "no_speech_prob": 1.4103068756224357e-12}, {"id": 61, "seek": 21118, "start": 216.4, "end": 218.66, "text": " and launch Claude Code as you normally would.", "tokens": [50626, 293, 4025, 12947, 2303, 15549, 382, 291, 5646, 576, 13, 50739], "temperature": 0, "avg_logprob": -0.15942821732486587, "compression_ratio": 1.5170731707317073, "no_speech_prob": 1.4103068756224357e-12}, {"id": 62, "seek": 21118, "start": 219.06, "end": 220.76000000000002, "text": " The harness never knows the difference.", "tokens": [50759, 440, 19700, 1128, 3255, 264, 2649, 13, 50844], "temperature": 0, "avg_logprob": -0.15942821732486587, "compression_ratio": 1.5170731707317073, "no_speech_prob": 1.4103068756224357e-12}, {"id": 63, "seek": 21118, "start": 221.22, "end": 224.46, "text": " It sends the same tool definitions, the same system prompt,", "tokens": [50867, 467, 14790, 264, 912, 2290, 21988, 11, 264, 912, 1185, 12391, 11, 51029], "temperature": 0, "avg_logprob": -0.15942821732486587, "compression_ratio": 1.5170731707317073, "no_speech_prob": 1.4103068756224357e-12}, {"id": 64, "seek": 21118, "start": 224.76000000000002, "end": 229.24, "text": " the same growing context, and K3 answers in the same shape Claude would.", "tokens": [51044, 264, 912, 4194, 4319, 11, 293, 591, 18, 6338, 294, 264, 912, 3909, 12947, 2303, 576, 13, 51268], "temperature": 0, "avg_logprob": -0.15942821732486587, "compression_ratio": 1.5170731707317073, "no_speech_prob": 1.4103068756224357e-12}, {"id": 65, "seek": 21118, "start": 229.66, "end": 231.38, "text": " Nothing else in your workflow changes.", "tokens": [51289, 6693, 1646, 294, 428, 20993, 2962, 13, 51375], "temperature": 0, "avg_logprob": -0.15942821732486587, "compression_ratio": 1.5170731707317073, "no_speech_prob": 1.4103068756224357e-12}, {"id": 66, "seek": 23138, "start": 231.38, "end": 234.84, "text": " If you use a router like OpenRouter, it is even shorter.", "tokens": [50365, 759, 291, 764, 257, 22492, 411, 7238, 49, 23985, 11, 309, 307, 754, 11639, 13, 50538], "temperature": 0, "avg_logprob": -0.13818284443446568, "compression_ratio": 1.510204081632653, "no_speech_prob": 1.4493977849514006e-12}, {"id": 67, "seek": 23138, "start": 235.38, "end": 239.01999999999998, "text": " K3 is already listed, priced exactly as Moonshot published,", "tokens": [50565, 591, 18, 307, 1217, 10052, 11, 30349, 2293, 382, 3335, 892, 12194, 6572, 11, 50747], "temperature": 0, "avg_logprob": -0.13818284443446568, "compression_ratio": 1.510204081632653, "no_speech_prob": 1.4493977849514006e-12}, {"id": 68, "seek": 23138, "start": 239.26, "end": 242.2, "text": " so you set one base URL, drop in your router key,", "tokens": [50759, 370, 291, 992, 472, 3096, 12905, 11, 3270, 294, 428, 22492, 2141, 11, 50906], "temperature": 0, "avg_logprob": -0.13818284443446568, "compression_ratio": 1.510204081632653, "no_speech_prob": 1.4493977849514006e-12}, {"id": 69, "seek": 23138, "start": 242.5, "end": 244.82, "text": " and you are running in a single config change.", "tokens": [50921, 293, 291, 366, 2614, 294, 257, 2167, 6662, 1319, 13, 51037], "temperature": 0, "avg_logprob": -0.13818284443446568, "compression_ratio": 1.510204081632653, "no_speech_prob": 1.4493977849514006e-12}, {"id": 70, "seek": 23138, "start": 245.14, "end": 248.8, "text": " And this is where free earns a real, if partial, meaning.", "tokens": [51053, 400, 341, 307, 689, 1737, 46936, 257, 957, 11, 498, 14641, 11, 3620, 13, 51236], "temperature": 0, "avg_logprob": -0.13818284443446568, "compression_ratio": 1.510204081632653, "no_speech_prob": 1.4493977849514006e-12}, {"id": 71, "seek": 23138, "start": 249.06, "end": 253.7, "text": " Some providers in the Kimi platform itself run promotional free tiers and trial credits at launch.", "tokens": [51249, 2188, 11330, 294, 264, 5652, 72, 3663, 2564, 1190, 41790, 1737, 40563, 293, 7308, 16816, 412, 4025, 13, 51481], "temperature": 0, "avg_logprob": -0.13818284443446568, "compression_ratio": 1.510204081632653, "no_speech_prob": 1.4493977849514006e-12}, {"id": 72, "seek": 25370, "start": 253.7, "end": 258.52, "text": " So for light use, you genuinely can drive Claude Code with K3 without paying for a while.", "tokens": [50365, 407, 337, 1442, 764, 11, 291, 17839, 393, 3332, 12947, 2303, 15549, 365, 591, 18, 1553, 6229, 337, 257, 1339, 13, 50606], "temperature": 0, "avg_logprob": -0.1087161086675689, "compression_ratio": 1.6458333333333333, "no_speech_prob": 1.367248220023043e-12}, {"id": 73, "seek": 25370, "start": 258.94, "end": 261.3, "text": " Not free forever, and not your own hardware,", "tokens": [50627, 1726, 1737, 5680, 11, 293, 406, 428, 1065, 8837, 11, 50745], "temperature": 0, "avg_logprob": -0.1087161086675689, "compression_ratio": 1.6458333333333333, "no_speech_prob": 1.367248220023043e-12}, {"id": 74, "seek": 25370, "start": 261.76, "end": 264.14, "text": " but the entry cost really can be zero to start.", "tokens": [50768, 457, 264, 8729, 2063, 534, 393, 312, 4018, 281, 722, 13, 50887], "temperature": 0, "avg_logprob": -0.1087161086675689, "compression_ratio": 1.6458333333333333, "no_speech_prob": 1.367248220023043e-12}, {"id": 75, "seek": 25370, "start": 264.53999999999996, "end": 266.76, "text": " That is the honest version of the free in the title.", "tokens": [50907, 663, 307, 264, 3245, 3037, 295, 264, 1737, 294, 264, 4876, 13, 51018], "temperature": 0, "avg_logprob": -0.1087161086675689, "compression_ratio": 1.6458333333333333, "no_speech_prob": 1.367248220023043e-12}, {"id": 76, "seek": 25370, "start": 267.18, "end": 269.56, "text": " One practical note before you switch everything over.", "tokens": [51039, 1485, 8496, 3637, 949, 291, 3679, 1203, 670, 13, 51158], "temperature": 0, "avg_logprob": -0.1087161086675689, "compression_ratio": 1.6458333333333333, "no_speech_prob": 1.367248220023043e-12}, {"id": 77, "seek": 25370, "start": 269.82, "end": 272.5, "text": " Test it on a real task first, not a toy prompt.", "tokens": [51171, 9279, 309, 322, 257, 957, 5633, 700, 11, 406, 257, 12058, 12391, 13, 51305], "temperature": 0, "avg_logprob": -0.1087161086675689, "compression_ratio": 1.6458333333333333, "no_speech_prob": 1.367248220023043e-12}, {"id": 78, "seek": 25370, "start": 272.82, "end": 277.12, "text": " Point K3 at an actual bug in an actual repo, and watch it run a full loop.", "tokens": [51321, 12387, 591, 18, 412, 364, 3539, 7426, 294, 364, 3539, 49040, 11, 293, 1159, 309, 1190, 257, 1577, 6367, 13, 51536], "temperature": 0, "avg_logprob": -0.1087161086675689, "compression_ratio": 1.6458333333333333, "no_speech_prob": 1.367248220023043e-12}, {"id": 79, "seek": 25370, "start": 277.34, "end": 279.56, "text": " What you are checking is not whether it can write a function.", "tokens": [51547, 708, 291, 366, 8568, 307, 406, 1968, 309, 393, 2464, 257, 2445, 13, 51658], "temperature": 0, "avg_logprob": -0.1087161086675689, "compression_ratio": 1.6458333333333333, "no_speech_prob": 1.367248220023043e-12}, {"id": 80, "seek": 27956, "start": 279.56, "end": 283.54, "text": " It is whether it holds the thread across 10 tool calls without losing the plan,", "tokens": [50365, 467, 307, 1968, 309, 9190, 264, 7207, 2108, 1266, 2290, 5498, 1553, 7027, 264, 1393, 11, 50564], "temperature": 0, "avg_logprob": -0.1720333993434906, "compression_ratio": 1.5714285714285714, "no_speech_prob": 1.6363663896123981e-12}, {"id": 81, "seek": 27956, "start": 283.74, "end": 285.5, "text": " and that only shows up on real work.", "tokens": [50574, 293, 300, 787, 3110, 493, 322, 957, 589, 13, 50662], "temperature": 0, "avg_logprob": -0.1720333993434906, "compression_ratio": 1.5714285714285714, "no_speech_prob": 1.6363663896123981e-12}, {"id": 82, "seek": 27956, "start": 285.82, "end": 287.1, "text": " And here is what that unlocks.", "tokens": [50678, 400, 510, 307, 437, 300, 517, 34896, 13, 50742], "temperature": 0, "avg_logprob": -0.1720333993434906, "compression_ratio": 1.5714285714285714, "no_speech_prob": 1.6363663896123981e-12}, {"id": 83, "seek": 27956, "start": 287.5, "end": 288.92, "text": " The token math flips hard.", "tokens": [50762, 440, 14862, 5221, 40249, 1152, 13, 50833], "temperature": 0, "avg_logprob": -0.1720333993434906, "compression_ratio": 1.5714285714285714, "no_speech_prob": 1.6363663896123981e-12}, {"id": 84, "seek": 27956, "start": 289.24, "end": 293.82, "text": " Kimi K3 runs at $3 per million input and $15 per million output.", "tokens": [50849, 5652, 72, 591, 18, 6676, 412, 1848, 18, 680, 2459, 4846, 293, 1848, 5211, 680, 2459, 5598, 13, 51078], "temperature": 0, "avg_logprob": -0.1720333993434906, "compression_ratio": 1.5714285714285714, "no_speech_prob": 1.6363663896123981e-12}, {"id": 85, "seek": 27956, "start": 294.58, "end": 296.82, "text": " Claude Fable 5 charges $10 and $50.", "tokens": [51116, 12947, 2303, 479, 712, 1025, 12235, 1848, 3279, 293, 1848, 2803, 13, 51228], "temperature": 0, "avg_logprob": -0.1720333993434906, "compression_ratio": 1.5714285714285714, "no_speech_prob": 1.6363663896123981e-12}, {"id": 86, "seek": 27956, "start": 297.36, "end": 300.96, "text": " So you are paying roughly a third for a model that lands three points behind.", "tokens": [51255, 407, 291, 366, 6229, 9810, 257, 2636, 337, 257, 2316, 300, 5949, 1045, 2793, 2261, 13, 51435], "temperature": 0, "avg_logprob": -0.1720333993434906, "compression_ratio": 1.5714285714285714, "no_speech_prob": 1.6363663896123981e-12}, {"id": 87, "seek": 27956, "start": 301.48, "end": 304.76, "text": " And if your workflow reuses context the way every coding agent does,", "tokens": [51461, 400, 498, 428, 20993, 319, 8355, 4319, 264, 636, 633, 17720, 9461, 775, 11, 51625], "temperature": 0, "avg_logprob": -0.1720333993434906, "compression_ratio": 1.5714285714285714, "no_speech_prob": 1.6363663896123981e-12}, {"id": 88, "seek": 27956, "start": 305.22, "end": 307.24, "text": " rereading the same files over and over,", "tokens": [51648, 46453, 8166, 264, 912, 7098, 670, 293, 670, 11, 51749], "temperature": 0, "avg_logprob": -0.1720333993434906, "compression_ratio": 1.5714285714285714, "no_speech_prob": 1.6363663896123981e-12}, {"id": 89, "seek": 30724, "start": 307.24, "end": 311.62, "text": " cached input drops to $0.30 a million, a tenth of the sticker price.", "tokens": [50365, 269, 15095, 4846, 11438, 281, 1848, 15, 13, 3446, 257, 2459, 11, 257, 27269, 295, 264, 20400, 3218, 13, 50584], "temperature": 0, "avg_logprob": -0.15005404608590261, "compression_ratio": 1.533112582781457, "no_speech_prob": 1.768513178783404e-12}, {"id": 90, "seek": 30724, "start": 312.0, "end": 313.02, "text": " Put it in real terms.", "tokens": [50603, 4935, 309, 294, 957, 2115, 13, 50654], "temperature": 0, "avg_logprob": -0.15005404608590261, "compression_ratio": 1.533112582781457, "no_speech_prob": 1.768513178783404e-12}, {"id": 91, "seek": 30724, "start": 313.36, "end": 317.90000000000003, "text": " A heavy day of agentic coding can burn 20 million tokens on Fable 5's output rate,", "tokens": [50671, 316, 4676, 786, 295, 9461, 299, 17720, 393, 5064, 945, 2459, 22667, 322, 479, 712, 1025, 311, 5598, 3314, 11, 50898], "temperature": 0, "avg_logprob": -0.15005404608590261, "compression_ratio": 1.533112582781457, "no_speech_prob": 1.768513178783404e-12}, {"id": 92, "seek": 30724, "start": 318.12, "end": 319.24, "text": " a bill that makes you flinch.", "tokens": [50909, 257, 2961, 300, 1669, 291, 932, 12415, 13, 50965], "temperature": 0, "avg_logprob": -0.15005404608590261, "compression_ratio": 1.533112582781457, "no_speech_prob": 1.768513178783404e-12}, {"id": 93, "seek": 30724, "start": 319.66, "end": 323.7, "text": " On K3, with caching on the input, the same work is closer to lunch money.", "tokens": [50986, 1282, 591, 18, 11, 365, 269, 2834, 322, 264, 4846, 11, 264, 912, 589, 307, 4966, 281, 6349, 1460, 13, 51188], "temperature": 0, "avg_logprob": -0.15005404608590261, "compression_ratio": 1.533112582781457, "no_speech_prob": 1.768513178783404e-12}, {"id": 94, "seek": 30724, "start": 324.16, "end": 327.2, "text": " Now, the word in the title I promise to be honest about.", "tokens": [51211, 823, 11, 264, 1349, 294, 264, 4876, 286, 6228, 281, 312, 3245, 466, 13, 51363], "temperature": 0, "avg_logprob": -0.15005404608590261, "compression_ratio": 1.533112582781457, "no_speech_prob": 1.768513178783404e-12}, {"id": 95, "seek": 30724, "start": 327.62, "end": 328.0, "text": " Local.", "tokens": [51384, 22755, 13, 51403], "temperature": 0, "avg_logprob": -0.15005404608590261, "compression_ratio": 1.533112582781457, "no_speech_prob": 1.768513178783404e-12}, {"id": 96, "seek": 30724, "start": 328.48, "end": 330.2, "text": " Because free and open source are true,", "tokens": [51427, 1436, 1737, 293, 1269, 4009, 366, 2074, 11, 51513], "temperature": 0, "avg_logprob": -0.15005404608590261, "compression_ratio": 1.533112582781457, "no_speech_prob": 1.768513178783404e-12}, {"id": 97, "seek": 30724, "start": 330.52, "end": 332.62, "text": " but local is where the hardware laughs at you.", "tokens": [51529, 457, 2654, 307, 689, 264, 8837, 6197, 412, 291, 13, 51634], "temperature": 0, "avg_logprob": -0.15005404608590261, "compression_ratio": 1.533112582781457, "no_speech_prob": 1.768513178783404e-12}, {"id": 98, "seek": 30724, "start": 333.02, "end": 335.96000000000004, "text": " Kimi K3 is 2.8 trillion parameters.", "tokens": [51654, 5652, 72, 591, 18, 307, 568, 13, 23, 18723, 9834, 13, 51801], "temperature": 0, "avg_logprob": -0.15005404608590261, "compression_ratio": 1.533112582781457, "no_speech_prob": 1.768513178783404e-12}, {"id": 99, "seek": 33596, "start": 335.96, "end": 338.46, "text": " At the 4-bit format Moonshot ships,", "tokens": [50365, 1711, 264, 1017, 12, 5260, 7877, 3335, 892, 12194, 11434, 11, 50490], "temperature": 0, "avg_logprob": -0.21380386352539063, "compression_ratio": 1.4190871369294606, "no_speech_prob": 2.3251546087815766e-12}, {"id": 100, "seek": 33596, "start": 338.91999999999996, "end": 342.14, "text": " that is about 1.4 terabytes just for the raw weights.", "tokens": [50513, 300, 307, 466, 502, 13, 19, 1796, 24538, 445, 337, 264, 8936, 17443, 13, 50674], "temperature": 0, "avg_logprob": -0.21380386352539063, "compression_ratio": 1.4190871369294606, "no_speech_prob": 2.3251546087815766e-12}, {"id": 101, "seek": 33596, "start": 342.64, "end": 346.26, "text": " The largest Mac studio tops out at 512 gigabytes.", "tokens": [50699, 440, 6443, 5707, 6811, 22836, 484, 412, 1025, 4762, 42741, 13, 50880], "temperature": 0, "avg_logprob": -0.21380386352539063, "compression_ratio": 1.4190871369294606, "no_speech_prob": 2.3251546087815766e-12}, {"id": 102, "seek": 33596, "start": 346.88, "end": 349.18, "text": " 4H100 give you 320.", "tokens": [50911, 1017, 39, 6879, 976, 291, 42429, 13, 51026], "temperature": 0, "avg_logprob": -0.21380386352539063, "compression_ratio": 1.4190871369294606, "no_speech_prob": 2.3251546087815766e-12}, {"id": 103, "seek": 33596, "start": 349.84, "end": 352.34, "text": " Even a single 8GPU B200 server,", "tokens": [51059, 2754, 257, 2167, 1649, 38, 8115, 363, 7629, 7154, 11, 51184], "temperature": 0, "avg_logprob": -0.21380386352539063, "compression_ratio": 1.4190871369294606, "no_speech_prob": 2.3251546087815766e-12}, {"id": 104, "seek": 33596, "start": 352.71999999999997, "end": 354.67999999999995, "text": " at 1.44 terabytes,", "tokens": [51203, 412, 502, 13, 13912, 1796, 24538, 11, 51301], "temperature": 0, "avg_logprob": -0.21380386352539063, "compression_ratio": 1.4190871369294606, "no_speech_prob": 2.3251546087815766e-12}, {"id": 105, "seek": 33596, "start": 355.21999999999997, "end": 358.26, "text": " barely clears the floor with nothing left for context or runtime.", "tokens": [51328, 10268, 47033, 264, 4123, 365, 1825, 1411, 337, 4319, 420, 34474, 13, 51480], "temperature": 0, "avg_logprob": -0.21380386352539063, "compression_ratio": 1.4190871369294606, "no_speech_prob": 2.3251546087815766e-12}, {"id": 106, "seek": 33596, "start": 358.79999999999995, "end": 362.29999999999995, "text": " Running the full K3 on your own machine is not a desktop project.", "tokens": [51507, 28136, 264, 1577, 591, 18, 322, 428, 1065, 3479, 307, 406, 257, 14502, 1716, 13, 51682], "temperature": 0, "avg_logprob": -0.21380386352539063, "compression_ratio": 1.4190871369294606, "no_speech_prob": 2.3251546087815766e-12}, {"id": 107, "seek": 36230, "start": 362.3, "end": 364.12, "text": " It is a data center project.", "tokens": [50365, 467, 307, 257, 1412, 3056, 1716, 13, 50456], "temperature": 0, "avg_logprob": -0.18855986228355995, "compression_ratio": 1.5104602510460252, "no_speech_prob": 2.8711507997492003e-12}, {"id": 108, "seek": 36230, "start": 364.44, "end": 365.92, "text": " So when a thumbnail says local,", "tokens": [50472, 407, 562, 257, 26746, 1619, 2654, 11, 50546], "temperature": 0, "avg_logprob": -0.18855986228355995, "compression_ratio": 1.5104602510460252, "no_speech_prob": 2.8711507997492003e-12}, {"id": 109, "seek": 36230, "start": 366.42, "end": 370.5, "text": " what it almost always means is running it through a provider on someone else's GPUs.", "tokens": [50571, 437, 309, 1920, 1009, 1355, 307, 2614, 309, 807, 257, 12398, 322, 1580, 1646, 311, 18407, 82, 13, 50775], "temperature": 0, "avg_logprob": -0.18855986228355995, "compression_ratio": 1.5104602510460252, "no_speech_prob": 2.8711507997492003e-12}, {"id": 110, "seek": 36230, "start": 370.72, "end": 371.6, "text": " The model is open.", "tokens": [50786, 440, 2316, 307, 1269, 13, 50830], "temperature": 0, "avg_logprob": -0.18855986228355995, "compression_ratio": 1.5104602510460252, "no_speech_prob": 2.8711507997492003e-12}, {"id": 111, "seek": 36230, "start": 372.02000000000004, "end": 373.56, "text": " Your laptop still cannot hold it.", "tokens": [50851, 2260, 10732, 920, 2644, 1797, 309, 13, 50928], "temperature": 0, "avg_logprob": -0.18855986228355995, "compression_ratio": 1.5104602510460252, "no_speech_prob": 2.8711507997492003e-12}, {"id": 112, "seek": 36230, "start": 373.98, "end": 375.18, "text": " Let me be fair, though,", "tokens": [50949, 961, 385, 312, 3143, 11, 1673, 11, 51009], "temperature": 0, "avg_logprob": -0.18855986228355995, "compression_ratio": 1.5104602510460252, "no_speech_prob": 2.8711507997492003e-12}, {"id": 113, "seek": 36230, "start": 375.46000000000004, "end": 377.34000000000003, "text": " because there is a real local story here,", "tokens": [51023, 570, 456, 307, 257, 957, 2654, 1657, 510, 11, 51117], "temperature": 0, "avg_logprob": -0.18855986228355995, "compression_ratio": 1.5104602510460252, "no_speech_prob": 2.8711507997492003e-12}, {"id": 114, "seek": 36230, "start": 377.54, "end": 378.44, "text": " just a smaller one.", "tokens": [51127, 445, 257, 4356, 472, 13, 51172], "temperature": 0, "avg_logprob": -0.18855986228355995, "compression_ratio": 1.5104602510460252, "no_speech_prob": 2.8711507997492003e-12}, {"id": 115, "seek": 36230, "start": 378.90000000000003, "end": 380.94, "text": " Open waits mean you are not locked to a vendor.", "tokens": [51195, 7238, 40597, 914, 291, 366, 406, 9376, 281, 257, 24321, 13, 51297], "temperature": 0, "avg_logprob": -0.18855986228355995, "compression_ratio": 1.5104602510460252, "no_speech_prob": 2.8711507997492003e-12}, {"id": 116, "seek": 36230, "start": 381.3, "end": 382.42, "text": " A company with the hardware,", "tokens": [51315, 316, 2237, 365, 264, 8837, 11, 51371], "temperature": 0, "avg_logprob": -0.18855986228355995, "compression_ratio": 1.5104602510460252, "no_speech_prob": 2.8711507997492003e-12}, {"id": 117, "seek": 38242, "start": 382.42, "end": 387.58000000000004, "text": " or a rented server genuinely can self-host K3 for privacy or compliance,", "tokens": [50365, 420, 257, 32381, 7154, 17839, 393, 2698, 12, 6037, 591, 18, 337, 11427, 420, 15882, 11, 50623], "temperature": 0, "avg_logprob": -0.15160835336107725, "compression_ratio": 1.5166051660516606, "no_speech_prob": 1.5607263745276434e-12}, {"id": 118, "seek": 38242, "start": 387.8, "end": 390.02000000000004, "text": " and no policy change can switch it off.", "tokens": [50634, 293, 572, 3897, 1319, 393, 3679, 309, 766, 13, 50745], "temperature": 0, "avg_logprob": -0.15160835336107725, "compression_ratio": 1.5166051660516606, "no_speech_prob": 1.5607263745276434e-12}, {"id": 119, "seek": 38242, "start": 390.34000000000003, "end": 390.7, "text": " Remember,", "tokens": [50761, 5459, 11, 50779], "temperature": 0, "avg_logprob": -0.15160835336107725, "compression_ratio": 1.5166051660516606, "no_speech_prob": 1.5607263745276434e-12}, {"id": 120, "seek": 38242, "start": 391.26, "end": 394.82, "text": " Fable 5 got pulled offline for days over an export control scare.", "tokens": [50807, 479, 712, 1025, 658, 7373, 21857, 337, 1708, 670, 364, 10725, 1969, 17185, 13, 50985], "temperature": 0, "avg_logprob": -0.15160835336107725, "compression_ratio": 1.5166051660516606, "no_speech_prob": 1.5607263745276434e-12}, {"id": 121, "seek": 38242, "start": 395.24, "end": 398.70000000000005, "text": " You cannot switch off a file already sitting on 10,000 hard drives.", "tokens": [51006, 509, 2644, 3679, 766, 257, 3991, 1217, 3798, 322, 1266, 11, 1360, 1152, 11754, 13, 51179], "temperature": 0, "avg_logprob": -0.15160835336107725, "compression_ratio": 1.5166051660516606, "no_speech_prob": 1.5607263745276434e-12}, {"id": 122, "seek": 38242, "start": 399.12, "end": 400.5, "text": " That durability is real,", "tokens": [51200, 663, 33664, 307, 957, 11, 51269], "temperature": 0, "avg_logprob": -0.15160835336107725, "compression_ratio": 1.5166051660516606, "no_speech_prob": 1.5607263745276434e-12}, {"id": 123, "seek": 38242, "start": 400.70000000000005, "end": 402.72, "text": " and it is the actual meaning of open here,", "tokens": [51279, 293, 309, 307, 264, 3539, 3620, 295, 1269, 510, 11, 51380], "temperature": 0, "avg_logprob": -0.15160835336107725, "compression_ratio": 1.5166051660516606, "no_speech_prob": 1.5607263745276434e-12}, {"id": 124, "seek": 38242, "start": 403.02000000000004, "end": 404.46000000000004, "text": " even if it never touches your desk.", "tokens": [51395, 754, 498, 309, 1128, 17431, 428, 10026, 13, 51467], "temperature": 0, "avg_logprob": -0.15160835336107725, "compression_ratio": 1.5166051660516606, "no_speech_prob": 1.5607263745276434e-12}, {"id": 125, "seek": 38242, "start": 404.56, "end": 407.0, "text": " And the honest downside does not stop at hardware.", "tokens": [51472, 400, 264, 3245, 25060, 775, 406, 1590, 412, 8837, 13, 51594], "temperature": 0, "avg_logprob": -0.15160835336107725, "compression_ratio": 1.5166051660516606, "no_speech_prob": 1.5607263745276434e-12}, {"id": 126, "seek": 40700, "start": 407.0, "end": 409.56, "text": " K3 trails the frontier on the hardest,", "tokens": [50365, 591, 18, 23024, 264, 35853, 322, 264, 13158, 11, 50493], "temperature": 0, "avg_logprob": -0.1849702254585598, "compression_ratio": 1.6080586080586081, "no_speech_prob": 2.193155379526246e-12}, {"id": 127, "seek": 40700, "start": 410.14, "end": 411.2, "text": " most novel problems,", "tokens": [50522, 881, 7613, 2740, 11, 50575], "temperature": 0, "avg_logprob": -0.1849702254585598, "compression_ratio": 1.6080586080586081, "no_speech_prob": 2.193155379526246e-12}, {"id": 128, "seek": 40700, "start": 411.44, "end": 414.9, "text": " and you will feel that gap on genuinely tricky architecture work.", "tokens": [50587, 293, 291, 486, 841, 300, 7417, 322, 17839, 12414, 9482, 589, 13, 50760], "temperature": 0, "avg_logprob": -0.1849702254585598, "compression_ratio": 1.6080586080586081, "no_speech_prob": 2.193155379526246e-12}, {"id": 129, "seek": 40700, "start": 415.28, "end": 418.12, "text": " Its safety and refusal behavior is different from Claude's,", "tokens": [50779, 6953, 4514, 293, 48948, 5223, 307, 819, 490, 12947, 2303, 311, 11, 50921], "temperature": 0, "avg_logprob": -0.1849702254585598, "compression_ratio": 1.6080586080586081, "no_speech_prob": 2.193155379526246e-12}, {"id": 130, "seek": 40700, "start": 418.4, "end": 420.64, "text": " tuned by a different lab with different defaults.", "tokens": [50935, 10870, 538, 257, 819, 2715, 365, 819, 7576, 82, 13, 51047], "temperature": 0, "avg_logprob": -0.1849702254585598, "compression_ratio": 1.6080586080586081, "no_speech_prob": 2.193155379526246e-12}, {"id": 131, "seek": 40700, "start": 420.82, "end": 424.04, "text": " And the polish of the Claude plus Claude code experience,", "tokens": [51056, 400, 264, 20452, 295, 264, 12947, 2303, 1804, 12947, 2303, 3089, 1752, 11, 51217], "temperature": 0, "avg_logprob": -0.1849702254585598, "compression_ratio": 1.6080586080586081, "no_speech_prob": 2.193155379526246e-12}, {"id": 132, "seek": 40700, "start": 424.5, "end": 427.38, "text": " the little ways the harness and the model were co-designed,", "tokens": [51240, 264, 707, 2098, 264, 19700, 293, 264, 2316, 645, 598, 12, 14792, 16690, 11, 51384], "temperature": 0, "avg_logprob": -0.1849702254585598, "compression_ratio": 1.6080586080586081, "no_speech_prob": 2.193155379526246e-12}, {"id": 133, "seek": 40700, "start": 427.82, "end": 430.88, "text": " is not something a swapped-in model perfectly reproduces.", "tokens": [51406, 307, 406, 746, 257, 50011, 12, 259, 2316, 6239, 11408, 887, 13, 51559], "temperature": 0, "avg_logprob": -0.1849702254585598, "compression_ratio": 1.6080586080586081, "no_speech_prob": 2.193155379526246e-12}, {"id": 134, "seek": 40700, "start": 431.22, "end": 431.96, "text": " It is close,", "tokens": [51576, 467, 307, 1998, 11, 51613], "temperature": 0, "avg_logprob": -0.1849702254585598, "compression_ratio": 1.6080586080586081, "no_speech_prob": 2.193155379526246e-12}, {"id": 135, "seek": 40700, "start": 432.48, "end": 433.16, "text": " not identical.", "tokens": [51639, 406, 14800, 13, 51673], "temperature": 0, "avg_logprob": -0.1849702254585598, "compression_ratio": 1.6080586080586081, "no_speech_prob": 2.193155379526246e-12}, {"id": 136, "seek": 43316, "start": 433.16, "end": 435.56, "text": " To be fair to just paying for Claude,", "tokens": [50365, 1407, 312, 3143, 281, 445, 6229, 337, 12947, 2303, 11, 50485], "temperature": 0, "avg_logprob": -0.19724610447883606, "compression_ratio": 1.5592592592592593, "no_speech_prob": 2.244349454547101e-12}, {"id": 137, "seek": 43316, "start": 435.94, "end": 437.92, "text": " that co-design is worth something.", "tokens": [50504, 300, 598, 12, 14792, 788, 307, 3163, 746, 13, 50603], "temperature": 0, "avg_logprob": -0.19724610447883606, "compression_ratio": 1.5592592592592593, "no_speech_prob": 2.244349454547101e-12}, {"id": 138, "seek": 43316, "start": 438.22, "end": 439.92, "text": " If your work is the hardest 10%,", "tokens": [50618, 759, 428, 589, 307, 264, 13158, 1266, 8923, 50703], "temperature": 0, "avg_logprob": -0.19724610447883606, "compression_ratio": 1.5592592592592593, "no_speech_prob": 2.244349454547101e-12}, {"id": 139, "seek": 43316, "start": 439.92, "end": 441.20000000000005, "text": " novel systems,", "tokens": [50703, 7613, 3652, 11, 50767], "temperature": 0, "avg_logprob": -0.19724610447883606, "compression_ratio": 1.5592592592592593, "no_speech_prob": 2.244349454547101e-12}, {"id": 140, "seek": 43316, "start": 441.72, "end": 442.96000000000004, "text": " subtle multi-file reasoning,", "tokens": [50793, 13743, 4825, 12, 69, 794, 21577, 11, 50855], "temperature": 0, "avg_logprob": -0.19724610447883606, "compression_ratio": 1.5592592592592593, "no_speech_prob": 2.244349454547101e-12}, {"id": 141, "seek": 43316, "start": 443.44, "end": 445.20000000000005, "text": " the frontier model earns its price,", "tokens": [50879, 264, 35853, 2316, 46936, 1080, 3218, 11, 50967], "temperature": 0, "avg_logprob": -0.19724610447883606, "compression_ratio": 1.5592592592592593, "no_speech_prob": 2.244349454547101e-12}, {"id": 142, "seek": 43316, "start": 445.6, "end": 449.08000000000004, "text": " and the smoothest experience is still Claude code pointed at Claude.", "tokens": [50987, 293, 264, 5508, 377, 1752, 307, 920, 12947, 2303, 3089, 10932, 412, 12947, 2303, 13, 51161], "temperature": 0, "avg_logprob": -0.19724610447883606, "compression_ratio": 1.5592592592592593, "no_speech_prob": 2.244349454547101e-12}, {"id": 143, "seek": 43316, "start": 449.46000000000004, "end": 451.84000000000003, "text": " Nobody optimized that pairing harder than Anthropic.", "tokens": [51180, 9297, 26941, 300, 32735, 6081, 813, 12727, 39173, 13, 51299], "temperature": 0, "avg_logprob": -0.19724610447883606, "compression_ratio": 1.5592592592592593, "no_speech_prob": 2.244349454547101e-12}, {"id": 144, "seek": 43316, "start": 452.20000000000005, "end": 453.76000000000005, "text": " So here is the honest verdict.", "tokens": [51317, 407, 510, 307, 264, 3245, 33957, 13, 51395], "temperature": 0, "avg_logprob": -0.19724610447883606, "compression_ratio": 1.5592592592592593, "no_speech_prob": 2.244349454547101e-12}, {"id": 145, "seek": 43316, "start": 454.0, "end": 455.74, "text": " If you run a lot of agentic coding,", "tokens": [51407, 759, 291, 1190, 257, 688, 295, 9461, 299, 17720, 11, 51494], "temperature": 0, "avg_logprob": -0.19724610447883606, "compression_ratio": 1.5592592592592593, "no_speech_prob": 2.244349454547101e-12}, {"id": 146, "seek": 43316, "start": 456.3, "end": 457.06, "text": " refactors,", "tokens": [51522, 1895, 578, 830, 11, 51560], "temperature": 0, "avg_logprob": -0.19724610447883606, "compression_ratio": 1.5592592592592593, "no_speech_prob": 2.244349454547101e-12}, {"id": 147, "seek": 43316, "start": 457.44000000000005, "end": 458.0, "text": " tests,", "tokens": [51579, 6921, 11, 51607], "temperature": 0, "avg_logprob": -0.19724610447883606, "compression_ratio": 1.5592592592592593, "no_speech_prob": 2.244349454547101e-12}, {"id": 148, "seek": 43316, "start": 458.5, "end": 459.12, "text": " glue code,", "tokens": [51632, 8998, 3089, 11, 51663], "temperature": 0, "avg_logprob": -0.19724610447883606, "compression_ratio": 1.5592592592592593, "no_speech_prob": 2.244349454547101e-12}, {"id": 149, "seek": 43316, "start": 459.54, "end": 461.06, "text": " the everyday 90%,", "tokens": [51684, 264, 7429, 4289, 8923, 51760], "temperature": 0, "avg_logprob": -0.19724610447883606, "compression_ratio": 1.5592592592592593, "no_speech_prob": 2.244349454547101e-12}, {"id": 150, "seek": 46106, "start": 461.06, "end": 464.68, "text": " Kimi K3 in Claude code is close to a cheat code.", "tokens": [50365, 5652, 72, 591, 18, 294, 12947, 2303, 3089, 307, 1998, 281, 257, 17470, 3089, 13, 50546], "temperature": 0, "avg_logprob": -0.19732495980669362, "compression_ratio": 1.5436241610738255, "no_speech_prob": 2.298399751771152e-12}, {"id": 151, "seek": 46106, "start": 465.1, "end": 466.42, "text": " You keep the harness you trust,", "tokens": [50567, 509, 1066, 264, 19700, 291, 3361, 11, 50633], "temperature": 0, "avg_logprob": -0.19732495980669362, "compression_ratio": 1.5436241610738255, "no_speech_prob": 2.298399751771152e-12}, {"id": 152, "seek": 46106, "start": 466.6, "end": 468.28000000000003, "text": " and cut your bill to a third or less.", "tokens": [50642, 293, 1723, 428, 2961, 281, 257, 2636, 420, 1570, 13, 50726], "temperature": 0, "avg_logprob": -0.19732495980669362, "compression_ratio": 1.5436241610738255, "no_speech_prob": 2.298399751771152e-12}, {"id": 153, "seek": 46106, "start": 468.78000000000003, "end": 470.14, "text": " Use K3 for volume.", "tokens": [50751, 8278, 591, 18, 337, 5523, 13, 50819], "temperature": 0, "avg_logprob": -0.19732495980669362, "compression_ratio": 1.5436241610738255, "no_speech_prob": 2.298399751771152e-12}, {"id": 154, "seek": 46106, "start": 470.64, "end": 474.36, "text": " Keep a frontier model one config flag away for the problems that actually need it.", "tokens": [50844, 5527, 257, 35853, 2316, 472, 6662, 7166, 1314, 337, 264, 2740, 300, 767, 643, 309, 13, 51030], "temperature": 0, "avg_logprob": -0.19732495980669362, "compression_ratio": 1.5436241610738255, "no_speech_prob": 2.298399751771152e-12}, {"id": 155, "seek": 46106, "start": 474.54, "end": 476.58, "text": " And treat local as run it through a provider,", "tokens": [51039, 400, 2387, 2654, 382, 1190, 309, 807, 257, 12398, 11, 51141], "temperature": 0, "avg_logprob": -0.19732495980669362, "compression_ratio": 1.5436241610738255, "no_speech_prob": 2.298399751771152e-12}, {"id": 156, "seek": 46106, "start": 476.78000000000003, "end": 478.9, "text": " unless you literally own a GPU server,", "tokens": [51151, 5969, 291, 3736, 1065, 257, 18407, 7154, 11, 51257], "temperature": 0, "avg_logprob": -0.19732495980669362, "compression_ratio": 1.5436241610738255, "no_speech_prob": 2.298399751771152e-12}, {"id": 157, "seek": 46106, "start": 479.26, "end": 481.36, "text": " because the full weights will never fit on a laptop.", "tokens": [51275, 570, 264, 1577, 17443, 486, 1128, 3318, 322, 257, 10732, 13, 51380], "temperature": 0, "avg_logprob": -0.19732495980669362, "compression_ratio": 1.5436241610738255, "no_speech_prob": 2.298399751771152e-12}, {"id": 158, "seek": 46106, "start": 481.74, "end": 483.34000000000003, "text": " Four numbers to walk away with.", "tokens": [51399, 7451, 3547, 281, 1792, 1314, 365, 13, 51479], "temperature": 0, "avg_logprob": -0.19732495980669362, "compression_ratio": 1.5436241610738255, "no_speech_prob": 2.298399751771152e-12}, {"id": 159, "seek": 46106, "start": 483.68, "end": 485.64, "text": " 2.8 trillion parameters.", "tokens": [51496, 568, 13, 23, 18723, 9834, 13, 51594], "temperature": 0, "avg_logprob": -0.19732495980669362, "compression_ratio": 1.5436241610738255, "no_speech_prob": 2.298399751771152e-12}, {"id": 160, "seek": 46106, "start": 485.84000000000003, "end": 488.32, "text": " Three points behind Fable 5 on intelligence.", "tokens": [51604, 6244, 2793, 2261, 479, 712, 1025, 322, 7599, 13, 51728], "temperature": 0, "avg_logprob": -0.19732495980669362, "compression_ratio": 1.5436241610738255, "no_speech_prob": 2.298399751771152e-12}, {"id": 161, "seek": 48832, "start": 488.32, "end": 489.86, "text": " Three dollars in,", "tokens": [50365, 6244, 3808, 294, 11, 50442], "temperature": 0, "avg_logprob": -0.18949792443252192, "compression_ratio": 1.5331010452961673, "no_speech_prob": 2.2190229254787086e-12}, {"id": 162, "seek": 48832, "start": 490.14, "end": 491.02, "text": " 15 out,", "tokens": [50456, 2119, 484, 11, 50500], "temperature": 0, "avg_logprob": -0.18949792443252192, "compression_ratio": 1.5331010452961673, "no_speech_prob": 2.2190229254787086e-12}, {"id": 163, "seek": 48832, "start": 491.34, "end": 492.74, "text": " a third of the frontier price,", "tokens": [50516, 257, 2636, 295, 264, 35853, 3218, 11, 50586], "temperature": 0, "avg_logprob": -0.18949792443252192, "compression_ratio": 1.5331010452961673, "no_speech_prob": 2.2190229254787086e-12}, {"id": 164, "seek": 48832, "start": 493.08, "end": 494.56, "text": " dropping to 30 cents cash.", "tokens": [50603, 13601, 281, 2217, 14941, 6388, 13, 50677], "temperature": 0, "avg_logprob": -0.18949792443252192, "compression_ratio": 1.5331010452961673, "no_speech_prob": 2.2190229254787086e-12}, {"id": 165, "seek": 48832, "start": 494.78, "end": 496.68, "text": " And 1.4 terabytes of weights,", "tokens": [50688, 400, 502, 13, 19, 1796, 24538, 295, 17443, 11, 50783], "temperature": 0, "avg_logprob": -0.18949792443252192, "compression_ratio": 1.5331010452961673, "no_speech_prob": 2.2190229254787086e-12}, {"id": 166, "seek": 48832, "start": 497.06, "end": 500.44, "text": " the number that decides whether local means your machine or someone else's.", "tokens": [50802, 264, 1230, 300, 14898, 1968, 2654, 1355, 428, 3479, 420, 1580, 1646, 311, 13, 50971], "temperature": 0, "avg_logprob": -0.18949792443252192, "compression_ratio": 1.5331010452961673, "no_speech_prob": 2.2190229254787086e-12}, {"id": 167, "seek": 48832, "start": 500.78, "end": 503.24, "text": " If this turned a hype thumbnail into an actual setup,", "tokens": [50988, 759, 341, 3574, 257, 24144, 26746, 666, 364, 3539, 8657, 11, 51111], "temperature": 0, "avg_logprob": -0.18949792443252192, "compression_ratio": 1.5331010452961673, "no_speech_prob": 2.2190229254787086e-12}, {"id": 168, "seek": 48832, "start": 503.48, "end": 505.18, "text": " you can run and a bill you can cut.", "tokens": [51123, 291, 393, 1190, 293, 257, 2961, 291, 393, 1723, 13, 51208], "temperature": 0, "avg_logprob": -0.18949792443252192, "compression_ratio": 1.5331010452961673, "no_speech_prob": 2.2190229254787086e-12}, {"id": 169, "seek": 48832, "start": 505.42, "end": 505.76, "text": " Subscribe.", "tokens": [51220, 10611, 13, 51237], "temperature": 0, "avg_logprob": -0.18949792443252192, "compression_ratio": 1.5331010452961673, "no_speech_prob": 2.2190229254787086e-12}, {"id": 170, "seek": 48832, "start": 506.34, "end": 510.92, "text": " This channel takes apart every open model and every local AI claim exactly like this.", "tokens": [51266, 639, 2269, 2516, 4936, 633, 1269, 2316, 293, 633, 2654, 7318, 3932, 2293, 411, 341, 13, 51495], "temperature": 0, "avg_logprob": -0.18949792443252192, "compression_ratio": 1.5331010452961673, "no_speech_prob": 2.2190229254787086e-12}, {"id": 171, "seek": 48832, "start": 511.18, "end": 512.86, "text": " Receipts first, verdict second.", "tokens": [51508, 41962, 48908, 700, 11, 33957, 1150, 13, 51592], "temperature": 0, "avg_logprob": -0.18949792443252192, "compression_ratio": 1.5331010452961673, "no_speech_prob": 2.2190229254787086e-12}, {"id": 172, "seek": 48832, "start": 512.86, "end": 514.54, "text": " I will see you in the next one.", "tokens": [51592, 286, 486, 536, 291, 294, 264, 958, 472, 13, 51676], "temperature": 0, "avg_logprob": -0.18949792443252192, "compression_ratio": 1.5331010452961673, "no_speech_prob": 2.2190229254787086e-12}], "language": "en"}