實際影片長度:8:35.000。原文、繁中、雙語可點擊句子跳轉影片。
0:00.000–0:04.541
You have seen the thumbnails by now. KimiK3 Plus Claude Code,
0:04.541–0:06.500
free, local, open source.
0:07.020–0:09.995
Point the best coding harness on the planet at an open model and
0:09.995–0:11.540
stop paying Anthropic forever.
0:12.060–0:16.051
It sounds too good, and the honest answer is that most of it is real,
0:16.051–0:17.720
and one word in that title is
0:17.720–0:20.525
quietly doing a lot of lifting, because
0:20.525–0:23.840
here is what is true. KimiK3 is genuinely open.
0:24.280–0:27.820
The weights are yours. It is a fraction of the price of the frontier,
0:27.820–0:29.978
and it does slot into Claude Code,
0:29.978–0:33.060
the exact agent loop most people already trust,
0:33.280–0:36.420
but the word local is where the fantasy and the hardware collide,
0:36.720–0:40.777
and I want to show you exactly where that line sits before you rip out your setup,
0:40.777–0:41.760
so here is the plan.
0:41.980–0:43.495
What KimiK3 actually is,
0:43.495–0:46.380
how you wire it into Claude Code in a few minutes,
0:46.760–0:50.820
and the honest truth about free versus local versus what your machine can really run.
0:51.200–0:52.589
Start with what shipped.
0:52.589–0:55.300
KimiK3 is an open-weight model from Moonshot AI,
0:55.300–0:58.220
released this month under a modified MIT license.
0:58.540–1:02.900
It is a mixture of experts model, 2.8 trillion parameters total,
1:03.320–1:07.560
but only 16 of its 896 experts fire for any given token,
1:07.940–1:11.020
so it is enormous on disk and lean when it actually runs.
1:11.420–1:13.872
It carries a 1 million token context window,
1:13.872–1:15.860
it reads images and video natively,
1:16.100–1:19.936
and its launch materials put it head-to-head with Claude Fable 5 on coding and
1:19.936–1:20.880
agent benchmarks.
1:21.180–1:22.140
That is the headline.
1:22.140–1:24.040
Here is what the headline leaves out.
1:24.480–1:27.940
On raw intelligence, K3 does not beat Fable 5.
1:28.400–1:34.040
Independent testing from artificial analysis puts it around 3 points behind the leader on the overall intelligence index,
1:34.260–1:36.680
but 3 points behind the best model on earth,
1:36.960–1:39.600
while being open and roughly a third of the price,
1:39.880–1:42.260
is a sentence that did not exist a month ago.
1:42.460–1:45.005
And on the axis that matters for a coding agent,
1:45.005–1:46.180
it is right in the mix.
1:46.180–1:51.340
It beats Opus 4.8 and GPT 5.5 on several coding benchmarks,
1:51.720–1:54.920
and wins 2 of 6 outright in Moonshot's own evaluation.
1:55.460–1:59.220
On GDP Vol, a benchmark of real economically valuable tasks,
1:59.520–2:02.000
it posts an ELO above Opus 4.8.
2:02.280–2:04.780
So the gap is real on the hardest novel problems,
2:05.020–2:08.240
and basically invisible on the everyday 90% of the job.
2:08.440–2:11.060
Now, why does this pair with Claude Code specifically?
2:11.060–2:13.340
Because Claude Code is just a harness.
2:13.740–2:15.260
Strip it down, and it is a loop.
2:15.580–2:18.240
Gather context, take an action through a tool call,
2:18.700–2:21.460
verify the result, repeat until the task is done.
2:21.880–2:25.640
The file edits, the shell commands, the plan and verify cycle,
2:26.000–2:27.320
all of it rides on that loop.
2:27.600–2:28.700
It was tuned for Claude,
2:29.020–2:31.620
but the loop does not actually care which model sits underneath,
2:32.040–2:33.900
as long as that model is good at two things,
2:34.340–2:35.820
calling tools in the right format,
2:36.240–2:37.960
and following instructions without drifting.
2:37.960–2:41.120
And K3 was built for exactly those two things.
2:41.480–2:43.800
Moonshot trained it for long horizon agent work,
2:44.140–2:47.000
the kind that runs for an hour instead of answering in one shot.
2:47.380–2:50.840
Read the repo, make a plan, edit 20 files,
2:51.380–2:53.320
run the tests, read the failures,
2:53.900–2:55.760
fix what broke, and try again.
2:56.120–2:58.392
That is the loop where most models drift and
2:58.392–2:59.780
fall apart around step 10.
3:00.100–3:02.560
It is also the exact loop Claude Code runs,
3:02.840–3:05.480
and the exact loop K3 was trained to survive.
3:05.480–3:08.820
There is even a K3 variant, K3 Swarm,
3:09.200–3:12.240
built to fan out many parallel copies on one big job,
3:12.560–3:15.980
the same shape as a coding agent spawning a dozen subtasks at once.
3:16.440–3:18.140
So how do you actually wire it up?
3:18.460–3:20.640
This is the part that is genuinely a few minutes.
3:21.200–3:23.540
Claude Code talks to an Anthropic-style endpoint,
3:24.000–3:26.380
and it reads that endpoint from environment variables.
3:26.800–3:30.291
So you point those variables at a provider serving Kimi K3 instead
3:30.291–3:31.180
of at Anthropic.
3:31.180–3:36.040
Set the base URL, set the API key, name the K3 model,
3:36.400–3:38.660
and launch Claude Code as you normally would.
3:39.060–3:40.760
The harness never knows the difference.
3:41.220–3:44.460
It sends the same tool definitions, the same system prompt,
3:44.760–3:46.627
the same growing context, and
3:46.627–3:49.240
K3 answers in the same shape Claude would.
3:49.660–3:51.380
Nothing else in your workflow changes.
3:51.380–3:54.840
If you use a router like OpenRouter, it is even shorter.
3:55.380–3:59.020
K3 is already listed, priced exactly as Moonshot published,
3:59.260–4:02.200
so you set one base URL, drop in your router key,
4:02.500–4:04.820
and you are running in a single config change.
4:05.140–4:08.800
And this is where free earns a real, if partial, meaning.
4:09.060–4:12.526
Some providers in the Kimi platform itself run promotional free tiers and
4:12.526–4:13.700
trial credits at launch.
4:13.700–4:14.624
So for light use,
4:14.624–4:15.615
you genuinely can
4:15.615–4:18.520
drive Claude Code with K3 without paying for a while.
4:18.940–4:21.300
Not free forever, and not your own hardware,
4:21.760–4:24.140
but the entry cost really can be zero to start.
4:24.540–4:26.760
That is the honest version of the free in the title.
4:27.180–4:29.560
One practical note before you switch everything over.
4:29.820–4:32.500
Test it on a real task first, not a toy prompt.
4:32.820–4:35.637
Point K3 at an actual bug in an actual repo, and
4:35.637–4:37.120
watch it run a full loop.
4:37.340–4:39.560
What you are checking is not whether it can write a function.
4:39.560–4:43.540
It is whether it holds the thread across 10 tool calls without losing the plan,
4:43.740–4:45.500
and that only shows up on real work.
4:45.820–4:47.100
And here is what that unlocks.
4:47.500–4:48.920
The token math flips hard.
4:49.240–4:53.820
Kimi K3 runs at $3 per million input and $15 per million output.
4:54.580–4:56.820
Claude Fable 5 charges $10 and $50.
4:57.360–5:00.960
So you are paying roughly a third for a model that lands three points behind.
5:01.480–5:04.760
And if your workflow reuses context the way every coding agent does,
5:05.220–5:07.240
rereading the same files over and over,
5:07.240–5:11.620
cached input drops to $0.30 a million, a tenth of the sticker price.
5:12.000–5:13.020
Put it in real terms.
5:13.360–5:17.900
A heavy day of agentic coding can burn 20 million tokens on Fable 5's output rate,
5:18.120–5:19.240
a bill that makes you flinch.
5:19.660–5:21.509
On K3, with caching on the input,
5:21.509–5:23.700
the same work is closer to lunch money.
5:24.160–5:27.200
Now, the word in the title I promise to be honest about.
5:27.620–5:28.000
Local.
5:28.480–5:30.200
Because free and open source are true,
5:30.520–5:32.620
but local is where the hardware laughs at you.
5:33.020–5:35.960
Kimi K3 is 2.8 trillion parameters.
5:35.960–5:38.460
At the 4-bit format Moonshot ships,
5:38.920–5:42.140
that is about 1.4 terabytes just for the raw weights.
5:42.640–5:46.260
The largest Mac studio tops out at 512 gigabytes.
5:46.880–5:49.180
4H100 give you 320.
5:49.840–5:52.340
Even a single 8GPU B200 server,
5:52.720–5:54.680
at 1.44 terabytes,
5:55.220–5:58.260
barely clears the floor with nothing left for context or runtime.
5:58.800–6:02.300
Running the full K3 on your own machine is not a desktop project.
6:02.300–6:04.120
It is a data center project.
6:04.440–6:05.920
So when a thumbnail says local,
6:06.420–6:10.500
what it almost always means is running it through a provider on someone else's GPUs.
6:10.720–6:11.600
The model is open.
6:12.020–6:13.560
Your laptop still cannot hold it.
6:13.980–6:15.180
Let me be fair, though,
6:15.460–6:17.340
because there is a real local story here,
6:17.540–6:18.440
just a smaller one.
6:18.900–6:20.940
Open waits mean you are not locked to a vendor.
6:21.300–6:22.420
A company with the hardware,
6:22.420–6:27.580
or a rented server genuinely can self-host K3 for privacy or compliance,
6:27.800–6:30.020
and no policy change can switch it off.
6:30.340–6:30.700
Remember,
6:31.260–6:34.820
Fable 5 got pulled offline for days over an export control scare.
6:35.240–6:38.700
You cannot switch off a file already sitting on 10,000 hard drives.
6:39.120–6:40.500
That durability is real,
6:40.700–6:42.720
and it is the actual meaning of open here,
6:43.020–6:44.460
even if it never touches your desk.
6:44.560–6:47.000
And the honest downside does not stop at hardware.
6:47.000–6:49.560
K3 trails the frontier on the hardest,
6:50.140–6:51.200
most novel problems,
6:51.440–6:54.900
and you will feel that gap on genuinely tricky architecture work.
6:55.280–6:58.120
Its safety and refusal behavior is different from Claude's,
6:58.400–7:00.640
tuned by a different lab with different defaults.
7:00.820–7:04.040
And the polish of the Claude plus Claude code experience,
7:04.500–7:07.380
the little ways the harness and the model were co-designed,
7:07.820–7:10.880
is not something a swapped-in model perfectly reproduces.
7:11.220–7:11.960
It is close,
7:12.480–7:13.160
not identical.
7:13.160–7:15.560
To be fair to just paying for Claude,
7:15.940–7:17.920
that co-design is worth something.
7:18.220–7:19.920
If your work is the hardest 10%,
7:19.920–7:21.200
novel systems,
7:21.720–7:22.960
subtle multi-file reasoning,
7:23.440–7:25.200
the frontier model earns its price,
7:25.600–7:29.080
and the smoothest experience is still Claude code pointed at Claude.
7:29.460–7:31.840
Nobody optimized that pairing harder than Anthropic.
7:32.200–7:33.760
So here is the honest verdict.
7:34.000–7:35.740
If you run a lot of agentic coding,
7:36.300–7:37.060
refactors,
7:37.440–7:38.000
tests,
7:38.500–7:39.120
glue code,
7:39.540–7:41.060
the everyday 90%,
7:41.060–7:44.680
Kimi K3 in Claude code is close to a cheat code.
7:45.100–7:46.420
You keep the harness you trust,
7:46.600–7:48.280
and cut your bill to a third or less.
7:48.780–7:50.140
Use K3 for volume.
7:50.640–7:54.360
Keep a frontier model one config flag away for the problems that actually need it.
7:54.540–7:56.580
And treat local as run it through a provider,
7:56.780–7:58.900
unless you literally own a GPU server,
7:59.260–8:01.360
because the full weights will never fit on a laptop.
8:01.740–8:03.340
Four numbers to walk away with.
8:03.680–8:05.640
2.8 trillion parameters.
8:05.840–8:08.320
Three points behind Fable 5 on intelligence.
8:08.320–8:09.860
Three dollars in,
8:10.140–8:11.020
15 out,
8:11.340–8:12.740
a third of the frontier price,
8:13.080–8:14.560
dropping to 30 cents cash.
8:14.780–8:16.680
And 1.4 terabytes of weights,
8:17.060–8:20.440
the number that decides whether local means your machine or someone else's.
8:20.780–8:23.240
If this turned a hype thumbnail into an actual setup,
8:23.480–8:25.180
you can run and a bill you can cut.
8:25.420–8:25.760
Subscribe.
8:26.340–8:28.791
This channel takes apart every open model and
8:28.791–8:30.920
every local AI claim exactly like this.
8:31.180–8:32.860
Receipts first, verdict second.
8:32.860–8:34.540
I will see you in the next one.
0:00.000–0:04.541
你現在應該已經看過縮圖了。KimiK3 Plus Claude Code,
0:04.541–0:06.500
免費、本地部署、開源。
0:07.020–0:09.995
將地球上最好的編碼工具指向開源模型,
0:09.995–0:11.540
並永遠停止向 Anthropic 付費。
0:12.060–0:16.051
這聽起來太好了,而誠實的答案是,其中大部分是真實的,
0:16.051–0:17.720
標題中的一個詞語
0:17.720–0:20.525
正在默默承擔大量工作,因為
0:20.525–0:23.840
事實如下。KimiK3 確實是開源的。
0:24.280–0:27.820
權重歸你所有。它的價格只是前沿模型的幾分之一,
0:27.820–0:29.978
並且它可以整合進 Claude Code,
0:29.978–0:33.060
這是大多數人已經信任的代理循環,
0:33.280–0:36.420
但「本地」一詞才是幻想與硬體碰撞的地方,
0:36.720–0:40.777
我想在你拆毀現有設定之前,向你展示這條界線究竟在哪裡,
0:40.777–0:41.760
所以這是計劃。
0:41.980–0:43.495
KimiK3 實際上是什麼,
0:43.495–0:46.380
如何在幾分鐘內將其接入 Claude Code,
0:46.760–0:50.820
以及關於免費、本地部署與你的機器真正能運行什麼的誠實真相。
0:51.200–0:52.589
從已發布的內容開始。
0:52.589–0:55.300
KimiK3 是 Moonshot AI 推出的開源權重模型,
0:55.300–0:58.220
本月以修改後的 MIT 許可證發布。
0:58.540–1:02.900
它是一個混合專家模型,總參數為 2.8 兆,
1:03.320–1:07.560
但對於任何給定的 token,其 896 個專家中只有 16 個被觸發,
1:07.940–1:11.020
因此它在磁碟上佔用空間巨大,但在實際運行時卻很輕量。
1:11.420–1:13.872
它擁有 100 萬 token 的上下文視窗,
1:13.872–1:15.860
原生支援讀取圖片和影片,
1:16.100–1:19.936
其發布資料顯示它在編碼和
1:19.936–1:20.880
代理基準測試中與 Claude Fable 5 正面對決。
1:21.180–1:22.140
這是頭條新聞。
1:22.140–1:24.040
以下是頭條新聞未提及的內容。
1:24.480–1:27.940
在原始智力方面,K3 並未擊敗 Fable 5。
1:28.400–1:34.040
來自 Artificial Analysis 的獨立測試顯示,它在整體智力指數上落後於領先者約 3 分,
1:34.260–1:36.680
但落後於地球上最好的模型 3 分,
1:36.960–1:39.600
同時又是開源的,且價格僅為其三分之一,
1:39.880–1:42.260
這是一個月前還不存在的一句話。
1:42.460–1:45.005
而在對編碼代理至關重要的軸線上,
1:45.005–1:46.180
它處於競爭行列中。
1:46.180–1:51.340
它在多項編碼基準測試中擊敗了 Opus 4.8 和 GPT 5.5,
1:51.720–1:54.920
並在 Moonshot 自己的評估中贏得了 6 項中的 2 項。
1:55.460–1:59.220
在 GDP Vol(一項衡量實際經濟價值任務的基準)上,
1:59.520–2:02.000
它取得了高於 Opus 4.8 的 ELO 評分。
2:02.280–2:04.780
所以在最困難的新穎問題上,差距是真實存在的,
2:05.020–2:08.240
但在日常工作中佔 90% 的部分,這種差距基本上看不出來。
2:08.440–2:11.060
那麼,為什麼這會特別搭配 Claude Code 使用呢?
2:11.060–2:13.340
因為 Claude Code 只是一個載具(harness)。
2:13.740–2:15.260
把它剝開來看,它其實是一個迴圈。
2:15.580–2:18.240
蒐集上下文,透過工具呼叫採取行動,
2:18.700–2:21.460
驗證結果,然後重複直到任務完成。
2:21.880–2:25.640
檔案編輯、Shell 指令、計畫與驗證的循環,
2:26.000–2:27.320
所有這些都依賴於那個迴圈。
2:27.600–2:28.700
它是為 Claude 調校過的,
2:29.020–2:31.620
但這個迴圈實際上並不關心底下運行的是哪個模型,
2:32.040–2:33.900
只要該模型擅長兩件事,
2:34.340–2:35.820
以正確的格式呼叫工具,
2:36.240–2:37.960
並且遵循指令而不偏離。
2:37.960–2:41.120
而 K3 正是為了這兩件事而打造的。
2:41.480–2:43.800
Moonshot 針對長程代理工作訓練了它,
2:44.140–2:47.000
那種需要運行一小時而不是單次回答的任務類型。
2:47.380–2:50.840
閱讀程式庫,制定計畫,編輯 20 個檔案,
2:51.380–2:53.320
執行測試,閱讀失敗訊息,
2:53.900–2:55.760
修復出錯的地方,然後再試一次。
2:56.120–2:58.392
這就是大多數模型會偏離並
2:58.392–2:59.780
在第 10 步左右崩潰的迴圈。
3:00.100–3:02.560
這也是 Claude Code 實際運行的迴圈,
3:02.840–3:05.480
以及 K3 被訓練來應付的迴圈。
3:05.480–3:08.820
甚至還有一個 K3 變體,K3 Swarm,
3:09.200–3:12.240
專門設計用來在單一大型任務中分派多個平行副本,
3:12.560–3:15.980
這與程式碼代理一次生成十幾個子任務的結構相同。
3:16.440–3:18.140
那麼,你實際上該如何連接它呢?
3:18.460–3:20.640
這部分是真正只需幾分鐘的事。
3:21.200–3:23.540
Claude Code 會連接到 Anthropic 風格的端點,
3:24.000–3:26.380
它會從環境變數中讀取該端點。
3:26.800–3:30.291
所以你要將這些變數指向提供 Kimi K3 的服務商,
3:30.291–3:31.180
而不是指向 Anthropic。
3:31.180–3:36.040
設定基礎 URL、設定 API 金鑰、命名 K3 模型,
3:36.400–3:38.660
然後像平常一樣啟動 Claude Code。
3:39.060–3:40.760
載具永遠不會察覺到差異。
3:41.220–3:44.460
它發送相同的工具定義、相同的系統提示,
3:44.760–3:46.627
相同的增長上下文,並且
3:46.627–3:49.240
K3 會以與 Claude 相同的格式回答。
3:49.660–3:51.380
你的工作流程中沒有其他任何改變。
3:51.380–3:54.840
如果你使用像 OpenRouter 這樣的路由器,流程甚至更短。
3:55.380–3:59.020
K3 已經被列出,定價完全按照 Moonshot 公佈的價格,
3:59.260–4:02.200
所以你只需設定一個基礎 URL,放入你的路由器金鑰,
4:02.500–4:04.820
然後只需單一的設定變更即可開始運行。
4:05.140–4:08.800
這正是「免費」一詞所具備的真實(儘管是部分)意義所在。
4:09.060–4:12.526
Kimi 平台本身的一些供應商在推出時會提供促銷免費方案
4:12.526–4:13.700
和試用額度。
4:13.700–4:14.624
因此,對於輕度使用,
4:14.624–4:15.615
你確實可以
4:15.615–4:18.520
在一段時間內不花錢的情況下,使用 K3 驅動 Claude Code。
4:18.940–4:21.300
這不是永遠免費,也不是使用你自己的硬體,
4:21.760–4:24.140
但起步成本確實可以是零。
4:24.540–4:26.760
這就是標題中「免費」一詞的誠實說法。
4:27.180–4:29.560
在切換一切之前,有一個實用建議。
4:29.820–4:32.500
先在真實任務上測試,而不是玩具提示。
4:32.820–4:35.637
讓 K3 針對實際儲存庫中的實際錯誤進行操作,並
4:35.637–4:37.120
觀察它完成整個循環。
4:37.340–4:39.560
你要檢查的不是它能否寫出一個函數。
4:39.560–4:43.540
而是它能否在 10 次工具呼叫中保持線索而不丟失計劃,
4:43.740–4:45.500
而這只有在真實工作中才會顯現。
4:45.820–4:47.100
這就是它所能解鎖的內容。
4:47.500–4:48.920
令牌的計算方式發生了巨大轉變。
4:49.240–4:53.820
Kimi K3 的輸入價格為每百萬個令牌 3 美元,輸出為每百萬個令牌 15 美元。
4:54.580–4:56.820
Claude Fable 5 則分別收取 10 美元和 50 美元。
4:57.360–5:00.960
因此,你支付的費用大約只有後者的三分之一,而該模型的性能卻低了三個等級。
5:01.480–5:04.760
如果你的工作流程像每個編碼代理那樣重複使用上下文,
5:05.220–5:07.240
反覆閱讀相同的文件,
5:07.240–5:11.620
快取的輸入成本降至每百萬個令牌 0.30 美元,僅為標價的十分之一。
5:12.000–5:13.020
讓我們用實際情況來說明。
5:13.360–5:17.900
在 Fable 5 的輸出速率下,一天重度代理編碼可能會消耗 2000 萬個令牌,
5:18.120–5:19.240
這筆帳單會讓你倒吸一口涼氣。
5:19.660–5:21.509
在 K3 上,並開啟輸入快取,
5:21.509–5:23.700
完成相同工作的成本僅相當於一頓午餐錢。
5:24.160–5:27.200
現在,來談談標題中我承諾要誠實面對的詞。
5:27.620–5:28.000
本地(Local)。
5:28.480–5:30.200
因為免費和開源是事實,
5:30.520–5:32.620
但「本地」才是讓硬體讓你哭笑不得的地方。
5:33.020–5:35.960
Kimi K3 擁有 2.8 兆參數。
5:35.960–5:38.460
在 Moonshot 提供的 4-bit 格式下,
5:38.920–5:42.140
僅原始權重就需要約 1.4 TB 的空間。
5:42.640–5:46.260
最大的 Mac Studio 最高僅有 512 GB。
5:46.880–5:49.180
H100 給您 320 GB。
5:49.840–5:52.340
即使是單台 8 GPU 的 B200 伺服器,
5:52.720–5:54.680
在 1.44 TB 的情況下,
5:55.220–5:58.260
勉強達到門檻,且沒有剩餘空間給上下文或執行階段使用。
5:58.800–6:02.300
在自己的機器上運行完整的 K3 並非桌面級專案。
6:02.300–6:04.120
這是一項資料中心級專案。
6:04.440–6:05.920
因此,當縮圖標示「本地端」時,
6:06.420–6:10.500
它幾乎總是意味著透過供應商在別人的 GPU 上運行。
6:10.720–6:11.600
該模型是開放的。
6:12.020–6:13.560
您的筆記型電腦仍然無法容納它。
6:13.980–6:15.180
不過讓我公平地說,
6:15.460–6:17.340
因為這裡確實有一個本地端的敘事,
6:17.540–6:18.440
只是規模較小。
6:18.900–6:20.940
開放等待意味著您不會被鎖定在單一供應商。
6:21.300–6:22.420
擁有硬體的公司,
6:22.420–6:27.580
或租用的伺服器,確實可以為了隱私或合規性而自行託管 K3,
6:27.800–6:30.020
且沒有任何政策變更可以將其關閉。
6:30.340–6:30.700
請記住,
6:31.260–6:34.820
《魔法門之英雄無敵 5》曾因出口管制恐慌而離線數天。
6:35.240–6:38.700
您無法關閉已經存放在 10,000 個硬碟上的檔案。
6:39.120–6:40.500
這種耐久性是真的,
6:40.700–6:42.720
這也是這裡「開放」的實際意義,
6:43.020–6:44.460
即使它從未出現在您的書桌上。
6:44.560–6:47.000
誠實來說,缺點並不僅限於硬體。
6:47.000–6:49.560
K3 在最困難、
6:50.140–6:51.200
最具創新性的問題上落後於前沿模型,
6:51.440–6:54.900
您會在真正棘手的架構工作中感受到這種差距。
6:55.280–6:58.120
它的安全性和拒絕行為與 Claude 不同,
6:58.400–7:00.640
由不同的實驗室以不同的預設值進行調整。
7:00.820–7:04.040
而 Claude 與 Claude Code 體驗的細膩度,
7:04.500–7:07.380
框架與模型共同設計的細微之處,
7:07.820–7:10.880
並非替換進來的模型能完美複製的。
7:11.220–7:11.960
它很接近,
7:12.480–7:13.160
但並非完全相同。
7:13.160–7:15.560
為了公平起見,僅支付 Claude 的費用,
7:15.940–7:17.920
這種共同設計的價值確實存在。
7:18.220–7:19.920
如果您的工作是難度最高的前 10%,
7:19.920–7:21.200
創新系統,
7:21.720–7:22.960
微妙的多檔案推理,
7:23.440–7:25.200
前沿模型才配得上它的價格,
7:25.600–7:29.080
而最順暢的體驗仍然是指向 Claude 的 Claude Code。
7:29.460–7:31.840
沒有人比 Anthropic 更致力於優化這組搭配。
7:32.200–7:33.760
所以這裡有個誠實的結論。
7:34.000–7:35.740
如果你進行大量的代理式程式設計,
7:36.300–7:37.060
重構,
7:37.440–7:38.000
測試,
7:38.500–7:39.120
粘合程式碼,
7:39.540–7:41.060
也就是日常那 90% 的工作,
7:41.060–7:44.680
在 Claude Code 中使用 Kimi K3 幾乎等於作弊碼。
7:45.100–7:46.420
保留你信任的框架,
7:46.600–7:48.280
並將費用砍到三分之一或更少。
7:48.780–7:50.140
用 K3 處理大量任務。
7:50.640–7:54.360
保留一個前沿模型,只需切換一個配置參數,即可應對真正需要它的問題。
7:54.540–7:56.580
並將「本地運行」視為透過供應商執行,
7:56.780–7:58.900
除非你實際上擁有一台 GPU 伺服器,
7:59.260–8:01.360
因為完整的權重永遠無法裝進筆記型電腦。
8:01.740–8:03.340
帶走四個關鍵數字。
8:03.680–8:05.640
2.8 兆參數。
8:05.840–8:08.320
在智慧程度上落後 Fable 5 三個點。
8:08.320–8:09.860
投入 3 美元,
8:10.140–8:11.020
產出 15 美元,
8:11.340–8:12.740
僅為前沿模型價格的三分之一,
8:13.080–8:14.560
現金價更降至 30 美分。
8:14.780–8:16.680
以及 1.4 TB 的權重,
8:17.060–8:20.440
這個數字決定了「本地運行」是指你的機器還是別人的機器。
8:20.780–8:23.240
如果這讓一個炒作性的縮圖變成了實際可行的方案,
8:23.480–8:25.180
你可以執行它,並削減你的帳單。
8:25.420–8:25.760
訂閱。
8:26.340–8:28.791
本頻道會像這樣拆解每一個開放模型,
8:28.791–8:30.920
以及每一個本地 AI 的聲明。
8:31.180–8:32.860
先出示證據,再給出結論。
8:32.860–8:34.540
我們下一部影片見。
0:00.000–0:04.541
You have seen the thumbnails by now. KimiK3 Plus Claude Code,
你現在應該已經看過縮圖了。KimiK3 Plus Claude Code,
0:04.541–0:06.500
free, local, open source.
免費、本地部署、開源。
0:07.020–0:09.995
Point the best coding harness on the planet at an open model and
將地球上最好的編碼工具指向開源模型,
0:09.995–0:11.540
stop paying Anthropic forever.
並永遠停止向 Anthropic 付費。
0:12.060–0:16.051
It sounds too good, and the honest answer is that most of it is real,
這聽起來太好了,而誠實的答案是,其中大部分是真實的,
0:16.051–0:17.720
and one word in that title is
標題中的一個詞語
0:17.720–0:20.525
quietly doing a lot of lifting, because
正在默默承擔大量工作,因為
0:20.525–0:23.840
here is what is true. KimiK3 is genuinely open.
事實如下。KimiK3 確實是開源的。
0:24.280–0:27.820
The weights are yours. It is a fraction of the price of the frontier,
權重歸你所有。它的價格只是前沿模型的幾分之一,
0:27.820–0:29.978
and it does slot into Claude Code,
並且它可以整合進 Claude Code,
0:29.978–0:33.060
the exact agent loop most people already trust,
這是大多數人已經信任的代理循環,
0:33.280–0:36.420
but the word local is where the fantasy and the hardware collide,
但「本地」一詞才是幻想與硬體碰撞的地方,
0:36.720–0:40.777
and I want to show you exactly where that line sits before you rip out your setup,
我想在你拆毀現有設定之前,向你展示這條界線究竟在哪裡,
0:40.777–0:41.760
so here is the plan.
所以這是計劃。
0:41.980–0:43.495
What KimiK3 actually is,
KimiK3 實際上是什麼,
0:43.495–0:46.380
how you wire it into Claude Code in a few minutes,
如何在幾分鐘內將其接入 Claude Code,
0:46.760–0:50.820
and the honest truth about free versus local versus what your machine can really run.
以及關於免費、本地部署與你的機器真正能運行什麼的誠實真相。
0:51.200–0:52.589
Start with what shipped.
從已發布的內容開始。
0:52.589–0:55.300
KimiK3 is an open-weight model from Moonshot AI,
KimiK3 是 Moonshot AI 推出的開源權重模型,
0:55.300–0:58.220
released this month under a modified MIT license.
本月以修改後的 MIT 許可證發布。
0:58.540–1:02.900
It is a mixture of experts model, 2.8 trillion parameters total,
它是一個混合專家模型,總參數為 2.8 兆,
1:03.320–1:07.560
but only 16 of its 896 experts fire for any given token,
但對於任何給定的 token,其 896 個專家中只有 16 個被觸發,
1:07.940–1:11.020
so it is enormous on disk and lean when it actually runs.
因此它在磁碟上佔用空間巨大,但在實際運行時卻很輕量。
1:11.420–1:13.872
It carries a 1 million token context window,
它擁有 100 萬 token 的上下文視窗,
1:13.872–1:15.860
it reads images and video natively,
原生支援讀取圖片和影片,
1:16.100–1:19.936
and its launch materials put it head-to-head with Claude Fable 5 on coding and
其發布資料顯示它在編碼和
1:19.936–1:20.880
agent benchmarks.
代理基準測試中與 Claude Fable 5 正面對決。
1:21.180–1:22.140
That is the headline.
這是頭條新聞。
1:22.140–1:24.040
Here is what the headline leaves out.
以下是頭條新聞未提及的內容。
1:24.480–1:27.940
On raw intelligence, K3 does not beat Fable 5.
在原始智力方面,K3 並未擊敗 Fable 5。
1:28.400–1:34.040
Independent testing from artificial analysis puts it around 3 points behind the leader on the overall intelligence index,
來自 Artificial Analysis 的獨立測試顯示,它在整體智力指數上落後於領先者約 3 分,
1:34.260–1:36.680
but 3 points behind the best model on earth,
但落後於地球上最好的模型 3 分,
1:36.960–1:39.600
while being open and roughly a third of the price,
同時又是開源的,且價格僅為其三分之一,
1:39.880–1:42.260
is a sentence that did not exist a month ago.
這是一個月前還不存在的一句話。
1:42.460–1:45.005
And on the axis that matters for a coding agent,
而在對編碼代理至關重要的軸線上,
1:45.005–1:46.180
it is right in the mix.
它處於競爭行列中。
1:46.180–1:51.340
It beats Opus 4.8 and GPT 5.5 on several coding benchmarks,
它在多項編碼基準測試中擊敗了 Opus 4.8 和 GPT 5.5,
1:51.720–1:54.920
and wins 2 of 6 outright in Moonshot's own evaluation.
並在 Moonshot 自己的評估中贏得了 6 項中的 2 項。
1:55.460–1:59.220
On GDP Vol, a benchmark of real economically valuable tasks,
在 GDP Vol(一項衡量實際經濟價值任務的基準)上,
1:59.520–2:02.000
it posts an ELO above Opus 4.8.
它取得了高於 Opus 4.8 的 ELO 評分。
2:02.280–2:04.780
So the gap is real on the hardest novel problems,
所以在最困難的新穎問題上,差距是真實存在的,
2:05.020–2:08.240
and basically invisible on the everyday 90% of the job.
但在日常工作中佔 90% 的部分,這種差距基本上看不出來。
2:08.440–2:11.060
Now, why does this pair with Claude Code specifically?
那麼,為什麼這會特別搭配 Claude Code 使用呢?
2:11.060–2:13.340
Because Claude Code is just a harness.
因為 Claude Code 只是一個載具(harness)。
2:13.740–2:15.260
Strip it down, and it is a loop.
把它剝開來看,它其實是一個迴圈。
2:15.580–2:18.240
Gather context, take an action through a tool call,
蒐集上下文,透過工具呼叫採取行動,
2:18.700–2:21.460
verify the result, repeat until the task is done.
驗證結果,然後重複直到任務完成。
2:21.880–2:25.640
The file edits, the shell commands, the plan and verify cycle,
檔案編輯、Shell 指令、計畫與驗證的循環,
2:26.000–2:27.320
all of it rides on that loop.
所有這些都依賴於那個迴圈。
2:27.600–2:28.700
It was tuned for Claude,
它是為 Claude 調校過的,
2:29.020–2:31.620
but the loop does not actually care which model sits underneath,
但這個迴圈實際上並不關心底下運行的是哪個模型,
2:32.040–2:33.900
as long as that model is good at two things,
只要該模型擅長兩件事,
2:34.340–2:35.820
calling tools in the right format,
以正確的格式呼叫工具,
2:36.240–2:37.960
and following instructions without drifting.
並且遵循指令而不偏離。
2:37.960–2:41.120
And K3 was built for exactly those two things.
而 K3 正是為了這兩件事而打造的。
2:41.480–2:43.800
Moonshot trained it for long horizon agent work,
Moonshot 針對長程代理工作訓練了它,
2:44.140–2:47.000
the kind that runs for an hour instead of answering in one shot.
那種需要運行一小時而不是單次回答的任務類型。
2:47.380–2:50.840
Read the repo, make a plan, edit 20 files,
閱讀程式庫,制定計畫,編輯 20 個檔案,
2:51.380–2:53.320
run the tests, read the failures,
執行測試,閱讀失敗訊息,
2:53.900–2:55.760
fix what broke, and try again.
修復出錯的地方,然後再試一次。
2:56.120–2:58.392
That is the loop where most models drift and
這就是大多數模型會偏離並
2:58.392–2:59.780
fall apart around step 10.
在第 10 步左右崩潰的迴圈。
3:00.100–3:02.560
It is also the exact loop Claude Code runs,
這也是 Claude Code 實際運行的迴圈,
3:02.840–3:05.480
and the exact loop K3 was trained to survive.
以及 K3 被訓練來應付的迴圈。
3:05.480–3:08.820
There is even a K3 variant, K3 Swarm,
甚至還有一個 K3 變體,K3 Swarm,
3:09.200–3:12.240
built to fan out many parallel copies on one big job,
專門設計用來在單一大型任務中分派多個平行副本,
3:12.560–3:15.980
the same shape as a coding agent spawning a dozen subtasks at once.
這與程式碼代理一次生成十幾個子任務的結構相同。
3:16.440–3:18.140
So how do you actually wire it up?
那麼,你實際上該如何連接它呢?
3:18.460–3:20.640
This is the part that is genuinely a few minutes.
這部分是真正只需幾分鐘的事。
3:21.200–3:23.540
Claude Code talks to an Anthropic-style endpoint,
Claude Code 會連接到 Anthropic 風格的端點,
3:24.000–3:26.380
and it reads that endpoint from environment variables.
它會從環境變數中讀取該端點。
3:26.800–3:30.291
So you point those variables at a provider serving Kimi K3 instead
所以你要將這些變數指向提供 Kimi K3 的服務商,
3:30.291–3:31.180
of at Anthropic.
而不是指向 Anthropic。
3:31.180–3:36.040
Set the base URL, set the API key, name the K3 model,
設定基礎 URL、設定 API 金鑰、命名 K3 模型,
3:36.400–3:38.660
and launch Claude Code as you normally would.
然後像平常一樣啟動 Claude Code。
3:39.060–3:40.760
The harness never knows the difference.
載具永遠不會察覺到差異。
3:41.220–3:44.460
It sends the same tool definitions, the same system prompt,
它發送相同的工具定義、相同的系統提示,
3:44.760–3:46.627
the same growing context, and
相同的增長上下文,並且
3:46.627–3:49.240
K3 answers in the same shape Claude would.
K3 會以與 Claude 相同的格式回答。
3:49.660–3:51.380
Nothing else in your workflow changes.
你的工作流程中沒有其他任何改變。
3:51.380–3:54.840
If you use a router like OpenRouter, it is even shorter.
如果你使用像 OpenRouter 這樣的路由器,流程甚至更短。
3:55.380–3:59.020
K3 is already listed, priced exactly as Moonshot published,
K3 已經被列出,定價完全按照 Moonshot 公佈的價格,
3:59.260–4:02.200
so you set one base URL, drop in your router key,
所以你只需設定一個基礎 URL,放入你的路由器金鑰,
4:02.500–4:04.820
and you are running in a single config change.
然後只需單一的設定變更即可開始運行。
4:05.140–4:08.800
And this is where free earns a real, if partial, meaning.
這正是「免費」一詞所具備的真實(儘管是部分)意義所在。
4:09.060–4:12.526
Some providers in the Kimi platform itself run promotional free tiers and
Kimi 平台本身的一些供應商在推出時會提供促銷免費方案
4:12.526–4:13.700
trial credits at launch.
和試用額度。
4:13.700–4:14.624
So for light use,
因此,對於輕度使用,
4:14.624–4:15.615
you genuinely can
你確實可以
4:15.615–4:18.520
drive Claude Code with K3 without paying for a while.
在一段時間內不花錢的情況下,使用 K3 驅動 Claude Code。
4:18.940–4:21.300
Not free forever, and not your own hardware,
這不是永遠免費,也不是使用你自己的硬體,
4:21.760–4:24.140
but the entry cost really can be zero to start.
但起步成本確實可以是零。
4:24.540–4:26.760
That is the honest version of the free in the title.
這就是標題中「免費」一詞的誠實說法。
4:27.180–4:29.560
One practical note before you switch everything over.
在切換一切之前,有一個實用建議。
4:29.820–4:32.500
Test it on a real task first, not a toy prompt.
先在真實任務上測試,而不是玩具提示。
4:32.820–4:35.637
Point K3 at an actual bug in an actual repo, and
讓 K3 針對實際儲存庫中的實際錯誤進行操作,並
4:35.637–4:37.120
watch it run a full loop.
觀察它完成整個循環。
4:37.340–4:39.560
What you are checking is not whether it can write a function.
你要檢查的不是它能否寫出一個函數。
4:39.560–4:43.540
It is whether it holds the thread across 10 tool calls without losing the plan,
而是它能否在 10 次工具呼叫中保持線索而不丟失計劃,
4:43.740–4:45.500
and that only shows up on real work.
而這只有在真實工作中才會顯現。
4:45.820–4:47.100
And here is what that unlocks.
這就是它所能解鎖的內容。
4:47.500–4:48.920
The token math flips hard.
令牌的計算方式發生了巨大轉變。
4:49.240–4:53.820
Kimi K3 runs at $3 per million input and $15 per million output.
Kimi K3 的輸入價格為每百萬個令牌 3 美元,輸出為每百萬個令牌 15 美元。
4:54.580–4:56.820
Claude Fable 5 charges $10 and $50.
Claude Fable 5 則分別收取 10 美元和 50 美元。
4:57.360–5:00.960
So you are paying roughly a third for a model that lands three points behind.
因此,你支付的費用大約只有後者的三分之一,而該模型的性能卻低了三個等級。
5:01.480–5:04.760
And if your workflow reuses context the way every coding agent does,
如果你的工作流程像每個編碼代理那樣重複使用上下文,
5:05.220–5:07.240
rereading the same files over and over,
反覆閱讀相同的文件,
5:07.240–5:11.620
cached input drops to $0.30 a million, a tenth of the sticker price.
快取的輸入成本降至每百萬個令牌 0.30 美元,僅為標價的十分之一。
5:12.000–5:13.020
Put it in real terms.
讓我們用實際情況來說明。
5:13.360–5:17.900
A heavy day of agentic coding can burn 20 million tokens on Fable 5's output rate,
在 Fable 5 的輸出速率下,一天重度代理編碼可能會消耗 2000 萬個令牌,
5:18.120–5:19.240
a bill that makes you flinch.
這筆帳單會讓你倒吸一口涼氣。
5:19.660–5:21.509
On K3, with caching on the input,
在 K3 上,並開啟輸入快取,
5:21.509–5:23.700
the same work is closer to lunch money.
完成相同工作的成本僅相當於一頓午餐錢。
5:24.160–5:27.200
Now, the word in the title I promise to be honest about.
現在,來談談標題中我承諾要誠實面對的詞。
5:27.620–5:28.000
Local.
本地(Local)。
5:28.480–5:30.200
Because free and open source are true,
因為免費和開源是事實,
5:30.520–5:32.620
but local is where the hardware laughs at you.
但「本地」才是讓硬體讓你哭笑不得的地方。
5:33.020–5:35.960
Kimi K3 is 2.8 trillion parameters.
Kimi K3 擁有 2.8 兆參數。
5:35.960–5:38.460
At the 4-bit format Moonshot ships,
在 Moonshot 提供的 4-bit 格式下,
5:38.920–5:42.140
that is about 1.4 terabytes just for the raw weights.
僅原始權重就需要約 1.4 TB 的空間。
5:42.640–5:46.260
The largest Mac studio tops out at 512 gigabytes.
最大的 Mac Studio 最高僅有 512 GB。
5:46.880–5:49.180
4H100 give you 320.
H100 給您 320 GB。
5:49.840–5:52.340
Even a single 8GPU B200 server,
即使是單台 8 GPU 的 B200 伺服器,
5:52.720–5:54.680
at 1.44 terabytes,
在 1.44 TB 的情況下,
5:55.220–5:58.260
barely clears the floor with nothing left for context or runtime.
勉強達到門檻,且沒有剩餘空間給上下文或執行階段使用。
5:58.800–6:02.300
Running the full K3 on your own machine is not a desktop project.
在自己的機器上運行完整的 K3 並非桌面級專案。
6:02.300–6:04.120
It is a data center project.
這是一項資料中心級專案。
6:04.440–6:05.920
So when a thumbnail says local,
因此,當縮圖標示「本地端」時,
6:06.420–6:10.500
what it almost always means is running it through a provider on someone else's GPUs.
它幾乎總是意味著透過供應商在別人的 GPU 上運行。
6:10.720–6:11.600
The model is open.
該模型是開放的。
6:12.020–6:13.560
Your laptop still cannot hold it.
您的筆記型電腦仍然無法容納它。
6:13.980–6:15.180
Let me be fair, though,
不過讓我公平地說,
6:15.460–6:17.340
because there is a real local story here,
因為這裡確實有一個本地端的敘事,
6:17.540–6:18.440
just a smaller one.
只是規模較小。
6:18.900–6:20.940
Open waits mean you are not locked to a vendor.
開放等待意味著您不會被鎖定在單一供應商。
6:21.300–6:22.420
A company with the hardware,
擁有硬體的公司,
6:22.420–6:27.580
or a rented server genuinely can self-host K3 for privacy or compliance,
或租用的伺服器,確實可以為了隱私或合規性而自行託管 K3,
6:27.800–6:30.020
and no policy change can switch it off.
且沒有任何政策變更可以將其關閉。
6:30.340–6:30.700
Remember,
請記住,
6:31.260–6:34.820
Fable 5 got pulled offline for days over an export control scare.
《魔法門之英雄無敵 5》曾因出口管制恐慌而離線數天。
6:35.240–6:38.700
You cannot switch off a file already sitting on 10,000 hard drives.
您無法關閉已經存放在 10,000 個硬碟上的檔案。
6:39.120–6:40.500
That durability is real,
這種耐久性是真的,
6:40.700–6:42.720
and it is the actual meaning of open here,
這也是這裡「開放」的實際意義,
6:43.020–6:44.460
even if it never touches your desk.
即使它從未出現在您的書桌上。
6:44.560–6:47.000
And the honest downside does not stop at hardware.
誠實來說,缺點並不僅限於硬體。
6:47.000–6:49.560
K3 trails the frontier on the hardest,
K3 在最困難、
6:50.140–6:51.200
most novel problems,
最具創新性的問題上落後於前沿模型,
6:51.440–6:54.900
and you will feel that gap on genuinely tricky architecture work.
您會在真正棘手的架構工作中感受到這種差距。
6:55.280–6:58.120
Its safety and refusal behavior is different from Claude's,
它的安全性和拒絕行為與 Claude 不同,
6:58.400–7:00.640
tuned by a different lab with different defaults.
由不同的實驗室以不同的預設值進行調整。
7:00.820–7:04.040
And the polish of the Claude plus Claude code experience,
而 Claude 與 Claude Code 體驗的細膩度,
7:04.500–7:07.380
the little ways the harness and the model were co-designed,
框架與模型共同設計的細微之處,
7:07.820–7:10.880
is not something a swapped-in model perfectly reproduces.
並非替換進來的模型能完美複製的。
7:11.220–7:11.960
It is close,
它很接近,
7:12.480–7:13.160
not identical.
但並非完全相同。
7:13.160–7:15.560
To be fair to just paying for Claude,
為了公平起見,僅支付 Claude 的費用,
7:15.940–7:17.920
that co-design is worth something.
這種共同設計的價值確實存在。
7:18.220–7:19.920
If your work is the hardest 10%,
如果您的工作是難度最高的前 10%,
7:19.920–7:21.200
novel systems,
創新系統,
7:21.720–7:22.960
subtle multi-file reasoning,
微妙的多檔案推理,
7:23.440–7:25.200
the frontier model earns its price,
前沿模型才配得上它的價格,
7:25.600–7:29.080
and the smoothest experience is still Claude code pointed at Claude.
而最順暢的體驗仍然是指向 Claude 的 Claude Code。
7:29.460–7:31.840
Nobody optimized that pairing harder than Anthropic.
沒有人比 Anthropic 更致力於優化這組搭配。
7:32.200–7:33.760
So here is the honest verdict.
所以這裡有個誠實的結論。
7:34.000–7:35.740
If you run a lot of agentic coding,
如果你進行大量的代理式程式設計,
7:36.300–7:37.060
refactors,
重構,
7:37.440–7:38.000
tests,
測試,
7:38.500–7:39.120
glue code,
粘合程式碼,
7:39.540–7:41.060
the everyday 90%,
也就是日常那 90% 的工作,
7:41.060–7:44.680
Kimi K3 in Claude code is close to a cheat code.
在 Claude Code 中使用 Kimi K3 幾乎等於作弊碼。
7:45.100–7:46.420
You keep the harness you trust,
保留你信任的框架,
7:46.600–7:48.280
and cut your bill to a third or less.
並將費用砍到三分之一或更少。
7:48.780–7:50.140
Use K3 for volume.
用 K3 處理大量任務。
7:50.640–7:54.360
Keep a frontier model one config flag away for the problems that actually need it.
保留一個前沿模型,只需切換一個配置參數,即可應對真正需要它的問題。
7:54.540–7:56.580
And treat local as run it through a provider,
並將「本地運行」視為透過供應商執行,
7:56.780–7:58.900
unless you literally own a GPU server,
除非你實際上擁有一台 GPU 伺服器,
7:59.260–8:01.360
because the full weights will never fit on a laptop.
因為完整的權重永遠無法裝進筆記型電腦。
8:01.740–8:03.340
Four numbers to walk away with.
帶走四個關鍵數字。
8:03.680–8:05.640
2.8 trillion parameters.
2.8 兆參數。
8:05.840–8:08.320
Three points behind Fable 5 on intelligence.
在智慧程度上落後 Fable 5 三個點。
8:08.320–8:09.860
Three dollars in,
投入 3 美元,
8:10.140–8:11.020
15 out,
產出 15 美元,
8:11.340–8:12.740
a third of the frontier price,
僅為前沿模型價格的三分之一,
8:13.080–8:14.560
dropping to 30 cents cash.
現金價更降至 30 美分。
8:14.780–8:16.680
And 1.4 terabytes of weights,
以及 1.4 TB 的權重,
8:17.060–8:20.440
the number that decides whether local means your machine or someone else's.
這個數字決定了「本地運行」是指你的機器還是別人的機器。
8:20.780–8:23.240
If this turned a hype thumbnail into an actual setup,
如果這讓一個炒作性的縮圖變成了實際可行的方案,
8:23.480–8:25.180
you can run and a bill you can cut.
你可以執行它,並削減你的帳單。
8:25.420–8:25.760
Subscribe.
訂閱。
8:26.340–8:28.791
This channel takes apart every open model and
本頻道會像這樣拆解每一個開放模型,
8:28.791–8:30.920
every local AI claim exactly like this.
以及每一個本地 AI 的聲明。
8:31.180–8:32.860
Receipts first, verdict second.
先出示證據,再給出結論。
8:32.860–8:34.540
I will see you in the next one.
我們下一部影片見。

影片筆記:Kimi K3 Is INSANE with Claude Code (FREE + Local + open Source)

一句話總結

影片解析 Moonshot AI 發布的開源模型 Kimi K3 如何透過 Claude Code 的代理循環(Agent Loop)進行整合,強調其在編碼任務上具備與前線模型競爭的能力,並透過快取機制大幅降低成本;同時釐清「免費」與「本地運行(Local)」的真實硬體限制,建議開發者將 K3 用於日常 90% 的編碼工作以節省費用,並保留前線模型處理高難度問題。

核心重點

  1. Kimi K3 模型規格與性能
  • 由 Moonshot AI 開發,採用修改版 MIT 授權。
  • 架構為混合專家模型(MoE),總參數 2.8 兆,每 token 僅激活 16 個專家(共 896 個),上下文視窗達 100 萬 token。
  • 在智力測試中略遜於前線模型(影片稱 Fable 5),但在編碼基準測試中擊敗 Opus 4.8 和 GPT 5.5。
  • 適合處理日常 90% 的編碼任務,但在最困難的新穎問題上仍有差距。
  1. 與 Claude Code 的整合機制
  • Claude Code 本質為代理循環(Agent Loop),不依賴特定模型,只要模型能正確格式化工具呼叫即可。
  • Kimi K3 針對長視角代理工作訓練,適合多步驟、反覆測試與修復的任務。
  • 透過設定環境變數(Base URL、API Key)即可將 Claude Code 指向提供 K3 的服務商(如 OpenRouter),工作流程與工具定義不變。
  1. 成本優勢與「免費」真相
  • K3 定價約為前線模型(影片稱 Fable 5)的三分之一。
  • 利用輸入快取(Cached input)可將成本進一步降至極低(每百萬 token $0.30)。
  • 「免費」僅限於部分供應商的促銷或試用信用額,並非永久免費。
  1. 「本地運行(Local)」的硬體現實
  • K3 在 4-bit 格式下權重約 1.4 TB。
  • 個人電腦(如 Mac Studio 512 GB、H100 320 GB)無法容納。
  • 單台 8-GPU B200 伺服器(約 1.44 TB)勉強達到門檻,無剩餘空間給上下文。
  • 標題中的「Local」實際多指透過供應商在雲端 GPU 上運行,開源的意義在於不受單一供應商綁定,可自託管以確保隱私或合規性。

詳細大綱

A. Kimi K3 模型本質與性能

  • 基本規格
  • 開發者:Moonshot AI。
  • 授權:修改版 MIT 授權(Modified MIT License)。
  • 架構:混合專家模型(Mixture of Experts, MoE)。
  • 參數總量:2.8 兆(2.8 trillion)。
  • 激活參數:每 token 僅激活 896 個專家中的 16 個。
  • 上下文視窗:100 萬 token。
  • 多模態能力:原生支援讀取圖片與影片。
  • 性能評估
  • 智力指數:在獨立測試(Artificial Analysis)中,整體智力指數比領先者(影片稱 Fable 5)低約 3 點。
  • 編碼表現
  • 在多個編碼基準測試中擊敗 Opus 4.8 和 GPT 5.5。
  • 在 Moonshot 自有的 6 項評估中贏得 2 項。
  • 在 GDP Vol(經濟價值任務基準)中,ELO 評分高於 Opus 4.8。
  • 適用場景:在處理日常 90% 的編碼任務時,差距幾乎不可見;但在最困難的新穎問題上,差距真實存在。

B. 與 Claude Code 的整合機制

  • 整合原理
  • Claude Code 本質上是一個代理循環(Agent Loop):收集上下文 -> 透過工具呼叫執行動作 -> 驗證結果 -> 重複直到任務完成。
  • 該循環不依賴特定模型,只要模型能正確格式化工具呼叫並遵循指令即可。
  • Kimi K3 針對長視角代理工作(Long horizon agent work)進行訓練,適合需要多步驟、反覆測試與修復的任務。
  • 存在變體 K3 Swarm,用於並行處理多個子任務,模擬編碼代理的行為。
  • 技術設定
  • Claude Code 透過環境變數讀取端點(Endpoint)。
  • 用戶只需設定 base URLAPI key 並指定 K3 模型名稱,即可指向提供 Kimi K3 的服務商(如 Anthropic 風格端點)。
  • 若使用路由器(如 OpenRouter),設定更簡便,僅需更改一個基礎 URL 和 API 金鑰。
  • 工作流程不變:相同的工具定義、系統提示詞、上下文增長,K3 以與 Claude 相同的格式回應。

C. 成本與「免費」的真實含義

  • 定價結構
  • Kimi K3:輸入每百萬 token $3,輸出每百萬 token $15。
  • Claude Opus 4(影片稱 Fable 5):輸入 $10,輸出 $50。
  • K3 價格約為前線模型的三分之一。
  • 快取優勢
  • 編碼代理會重複讀取相同檔案,輸入快取(Cached input)可將成本降至每百萬 token $0.30。
  • 對比:重度代理編碼日在 Opus 4 下可能產生高昂費用,而在 K3 下成本極低。
  • 「免費」的定義
  • 部分供應商(Kimi 平台自身)提供促銷免費層級或試用信用額。
  • 輕度使用可暫時不付費,但並非「永遠免費」,也非「本地硬體免費」。

D. 「本地運行(Local)」的硬體現實

  • 硬體需求
  • Kimi K3 在 4-bit 格式下的權重大約為 1.4 TB
  • Mac Studio:最大僅 512 GB,無法容納。
  • H100:僅 320 GB。
  • 單台 8-GPU B200 伺服器:約 1.44 TB,勉強達到門檻,無剩餘空間給上下文或運行時。
  • 結論
  • 在個人電腦上運行完整 K3 不是桌面專案,而是資料中心專案。
  • 標題中的「Local」實際意義通常是透過供應商在別人的 GPU 上運行。
  • 開源的真正意義:權重開放意味著不受單一供應商綁定,企業可自託管以確保隱私或合規性,且不會因政策變更(如出口管制)而離線。

E. 缺點與最終建議

  • 缺點
  • 在最困難、最新穎的問題上落後於前線模型。
  • 安全與拒絕行為(Safety and refusal behavior)與 Claude 不同,由不同實驗室訓練。
  • 缺乏 Claude 與 Claude Code 之間經過共同設計的細微優化(Co-design)。
  • 最終建議
  • 日常使用:對於代理編碼、重構、測試、粘合代碼(Glue code)等日常 90% 的工作,Kimi K3 在 Claude Code 中接近「作弊碼」。
  • 策略:使用 K3 處理大量工作以節省費用;保留一個前線模型(透過一個配置標誌切換)以處理真正需要高智力的問題。
  • 本地定義:除非擁有 GPU 伺服器,否則「本地」應理解為透過供應商運行。

工具 / 模型 / 名詞整理

  • 模型
  • Kimi K3(含變體 K3 Swarm)
  • Claude Opus 4(逐字稿中多次誤植為 Claude Fable 5Fable 5
  • GPT 5.5
  • 工具/平台
  • Claude Code
  • Moonshot AI(開發者)
  • OpenRouter(路由器服務)
  • Artificial Analysis(獨立測試機構)
  • 硬體/規格
  • Mac Studio(提及最大記憶體 512 GB)
  • H100(提及顯存 320 GB)
  • B200(提及 8-GPU 伺服器)
  • 其他專有名詞
  • MIT License(修改版)
  • MoE(Mixture of Experts,混合專家模型)
  • GDP Vol(經濟價值任務基準測試)
  • ELO(評分系統)

操作流程整理

  1. 準備環境
  • 確保擁有 Claude Code 環境。
  • 獲取提供 Kimi K3 的服務商端點資訊(Base URL)與 API Key(例如透過 OpenRouter)。
  1. 配置 Claude Code
  • 設定環境變數,指向 Kimi K3 的端點。
  • 指定模型名稱為 K3。
  1. 執行代理任務
  • 啟動 Claude Code,其代理循環(Agent Loop)開始運作。
  • K3 接收上下文,格式化工具呼叫,執行動作並驗證結果。
  • 重複上述步驟直到任務完成。
  1. 監控成本與效能
  • 利用輸入快取(Cached input)降低重複讀取檔案的成本。
  • 觀察 K3 在編碼任務上的表現,若遇到高難度問題,可切換回前線模型。

值得注意的限制或風險

  1. 硬體限制
  • Kimi K3 權重約 1.4 TB,個人電腦(Mac Studio, H100)無法本地運行完整模型。
  • 單台 8-GPU B200 伺服器勉強容納,無剩餘空間給上下文或運行時。
  1. 效能差距
  • 在智力測試中略遜於前線模型(影片稱 Fable 5)。
  • 在最困難、最新穎的問題上,K3 落後於前線模型。
  1. 安全與行為差異
  • K3 的安全與拒絕行為與 Claude 不同,由不同實驗室訓練。
  • 缺乏 Claude 與 Claude Code 之間經過共同設計的細微優化。
  1. 「免費」誤解
  • 「免費」僅限於部分供應商的促銷或試用信用額,並非永久免費。
  • 「本地運行」在個人硬體層面不現實,實際多為雲端供應商運行。

逐字稿辨識疑點

  • Claude Fable 5 / Fable 5:逐字稿中多次將 Anthropic 的頂級模型稱為 "Claude Fable 5" 或 "Fable 5",並提及 "Claude Opus 4.8" 和 "GPT 5.5"。根據當前市場資訊,Anthropic 的頂級模型為 Opus,OpenAI 為 GPT-4o 或 GPT-4.5 等,"Fable 5" 與 "Opus 4.8" 及 "GPT 5.5" 均與已知產品名稱不符,疑為口誤或聽寫錯誤。
  • 4H100:逐字稿中提到 "4H100 give you 320",疑為 "H100" 的聽寫錯誤或特定硬體型號誤植。
  • GDP Vol:作為基準測試名稱,需查證是否為 "GPQA" 或其他類似名稱的誤植,或為特定內部測試名稱。
  • KimiK3:逐字稿開頭使用 "KimiK3",後文使用 "Kimi K3" 或 "K3",需確認官方命名是否為連寫。
  • 1.44 terabytes:逐字稿提到單台 8-GPU B200 伺服器僅有 1.44 TB 空間來容納 1.4 TB 權重,需查證 B200 的顯存總量是否確實如此低(通常 B200 單卡顯存為 192GB 或更高,8卡總和應遠高於此數值,此處數據可能為口誤或特定配置誤解)。

可延伸追問

  1. Kimi K3 的「修改版 MIT 授權」具體限制了哪些商業使用場景?
  2. 如何具體配置 OpenRouter 或其他路由器以最佳化 Claude Code 對 K3 的呼叫?
  3. 在什麼情況下,K3 的「安全與拒絕行為」差異會對開發者造成實際困擾?
  4. 若企業希望真正「本地運行」Kimi K3,需要具備怎樣的硬體基礎設施與成本效益分析?
  5. Kimi K3 的「K3 Swarm」變體在實際編碼代理任務中如何具體運作?

尚未產生學習筆記

請在 Telegram 指令最後加上「學習」,例如:videonote 網址 英文 雙語 學習