{"text": "昨天晚上,月資暗面把KIMI K3正式開放了。這個模型幹了兩件非常有衝擊力的事情。第一,他把全球開源模型的參數記錄直接翻了一個倍。K3模型的參數是2.8萬億。什麼概念呢?在之前的開源陣營裡面,最大的是Deepseek,只有1.6萬億。K3的大小差不多是他的兩倍。第二,在一項衡量真實知識工作的評測裡面,他排到了全球第三。第一名和第二名分別是Cloud5和GPT5.6瘦。而且要注意的是,Cloud是OPUS4.8被他擠到了後面一名。開源模型第一次站到了B原刺旗艦模型的前面。那麼這期視頻呢,我們就來詳細的了解一下KIMI K3,它到底強在哪?價格又有多恒?先來看一下時間線。其實在7月15號,KIMI的開放平台的一個促銷頁面已經提前泄露了。大家就知道可能有新的模型要上了。那麼在16號,也就是昨天的晚上,KIMI K3正式上線。分別在KIMI的APP以及KIMI Code裡面都可以使用了。現在只要進入網站APP,包括如果你是KIMI Code的會員,都可以正式使用KIMI K3了。這個時間點呢,也非常的微妙。正好卡在上海的世界人工智能大會前面。而且呢,月之岸面也正在談新的一輪融資。估值呢,大約是315億美元。KIMI K3這個成績單是給誰看的,就不用我多說了。然後來看一下啊,規格。2.8萬億這個參數呢,肯定是非常吸引眼球的。但是我覺得比大更值得看的,是他怎麼樣把這麼大的模型給他跑起來。看官方放的這張結構圖呢,我們也可以看到有兩個新的東西。一個叫KDA,KIMI Delta Tension混合線性注意力。簡單解釋就是,傳統注意力的計算量呢,會隨著上下文長度平方級上漲,越長呢就越貴。那麼K3的做法是呢,是每四層裡面,三層用便宜的線性注意力。然後一層用完整的全注意力。這裡省下來的算力,換成一個東西,100萬Token的上下文。之前KIMI K2的上下文長度呢,只有256K,直接翻了4倍。另外一個新的東西叫注意力殘差。讓深層網絡把淺層的注意力信息呢,直接接力上來。官方說,這套組合讓整體的訓練效率做到了KIMI K2的2.5倍。然後就是混合專家系統這塊,做的也非常非常的激進。整個模型呢,有896個專家,每次推理呢只激活16個。也就是說,雖然總參數量是2.8萬億,但是真正幹活的每次呢,只是裡面很小的一部分。再來看一下成績單,我挑三個重點的在視頻裡面說一下。第一個呢,是剛才視頻開頭提到的GPTWALL,他測的呢不是做題,是44個職業裡面真實的工作任務。那麼K3呢,是拿了1687分,全球第三名。第二個呢,是GP2A Diamond,研究生級別的科學題,他的得分呢是93.5%,是目前所有看源模型裡面得分最高的。第三個呢,我覺得是最實在的BlowComp,在網上檢索問題的答案,得分呢是91.2%,也是當時公開成績裡面最高的。那麼這一下呢,很明顯是吃到了100萬上下文的紅利。當然,測評呢也只能說明一半。另一半呢,我們還要真實上手去測試。這裡呢,我測試了兩個案例。第一個測試呢,是讓它根據官方的報導,以及我自己整理的文件。讓它直接幫我生成一份Kimi K3不低於12頁的一個PPT,幻燈篇,並且直接以網頁的形式展現。那麼這個呢,就是它一次性完成的任務。可以先從第一屏看一下,Kimi K3,它用模型第一次摸到了B原模型次旗艦的門檻。這是等於是PPT的封面。然後它最主要的幾個參數,它的配色,它的佈局,還有這些數據的提取,全部是Kimi K3的,自動去完成的。我覺得這個版面的佈局,以及參數的提取,配色其實都做得非常好。而且內容呢非常的詳盡,非常的豐富。還有關於它的這個架構拆解,這種圖表設計,也全部是它一次性完成的。目前看起來沒有找到很明顯的錯誤。基本上可以達到一個直接使用的水平了。而且是沒有創建任何Scale的前提下,只是給它發了一個簡單的報告,以及一個Planter,讓它生成PPT,就完成了。這能力我覺得還是不錯的。然後我在測試過程當中呢,發現一個比較明顯的問題。就是不算是現在使用的人比較多還是什麼的。就是我發送了這樣一條Planter過去,它整個完成這個任務的耗時,時間非常的長。至少花費了十幾分鐘。那這個任務如果以往是用GPT或者用Cloud去完成,我覺得這時間肯定是在10分鐘以內的。因為任務並不是特別的難。然後它的內容也不是很多。用KimiK3生成任務的時間是超出了我的預期的。然後第二個任務呢,是我讓它基於我前面兩期視頻。我是創建了一個體感遊戲。用谷歌的模型,它可以監控我們人體的關節、部位。首先我是做了一個體感監視器。這個在上期視頻已經講過了。包括它的項目,我已經開運到KitaHub上面去了。後面呢,我又用GPT5.64添加了幾個體感遊戲。比如有體感穿墻、星球互聯、還有這個切水果的遊戲,以及雙輪對打的遊戲。這個是之前評測Fibre5以及GPT5.6生成的。然後今天呢,我是想用這個KimiK3去做一個體感消消樂的遊戲。那其實像這套框架,在我原來的項目裡面,它已經生成了。也就是它的模型、它的底層邏輯、框架都有了。只是讓它在這個基礎上去增加一個體感消消樂的遊戲。那麼今天用KimiK3出現了兩個問題。第一個問題呢,就是剛才我提到的。它的耗時呢,非常非常的久。運行了差不多十幾二十分鐘,都沒有完成。然後第二個問題呢,就是我目前開通的是Kimi的第二個等級的套餐。也就是一個月是99塊錢。那麼是可以在KimiCode裡面去使用這個KimiK3的模型。如果你開通的是第一個等級的套餐。目前在KimiCode裡面是使用不了這個K3的模型的。完事呢,我在測試第二個任務的時候,可以看到,它給我羅列的toDo裡面有這麼多步驟。我覺得這個步驟是設計的有點過於冷餘的。然後在做第二個任務的時候,僅僅只是完成了第一條toDo,它5個小時的用量限制就已經達到了。然後並且呢,週用量也達到了20%的使用。也就是說,我開通99塊錢一個月的會員,僅僅只是完成了一個多的任務。第二個任務還沒有完成。有可能是因為這個模型太大,然後現在用的人太多。所以它整體的一個使用量限制比較少。包括我開通的這個第二等級的會員,它並不支持100萬的上下文。只有256K,那麼你必須是開通這個200塊錢一個月的,才可以享受100萬的上下文長度。給我的感覺還是不知道因為算力,還是什麼其他原因啊。誠意不是特別的夠。接下來呢,說一下價格。這個可能是對使用者、開發者殺傷力最大的部分。輸入呢,是3美元每100萬Token,輸出呢是15美元。如果是緩存命中的話,輸入只要3毛。這一塊呢,一直是KIMI做的比較好的地方。OK,那麼強的地方呢,已經說完了。接下來呢,來潑兩盆冷水。第一盤呢,全球最大開用模型的這個說法呢,它現在還只是一個期貨。模型能用是真的,但是呢,權重還沒有開放。官方是承諾在7月27號之前,放出完整的權重和技術報告。第二盤呢,就算是權重開放了,我們自己本地電腦呢,也是跑不動的。2.8萬億參數呢,就算你有1.5TB內存的MGDODO集群,跑起來都夠欠。這是一個數據中心級別的模型。所以,開源對於個人的意義呢,更多是讓第三方平台能便宜的把它跑起來,而不是說把它下載到本地自己去跑。", "segments": [{"id": 0, "seek": 0, "start": 0.0, "end": 5.38, "text": "昨天晚上,月資暗面把KIMI K3正式開放了。", "tokens": [50365, 47404, 6135, 50157, 11, 6939, 35380, 23826, 245, 8833, 16075, 42, 6324, 40, 591, 18, 15789, 27584, 8949, 12744, 2289, 1543, 50634], "temperature": 0, "avg_logprob": -0.23796453147098937, "compression_ratio": 1.1478599221789882, "no_speech_prob": 6.594049958841275e-12}, {"id": 1, "seek": 0, "start": 5.64, "end": 9.98, "text": "這個模型幹了兩件非常有衝擊力的事情。", "tokens": [50647, 6287, 41908, 39823, 29932, 2289, 16313, 20485, 14392, 2412, 9890, 251, 37813, 13486, 1546, 24675, 1543, 50864], "temperature": 0, "avg_logprob": -0.23796453147098937, "compression_ratio": 1.1478599221789882, "no_speech_prob": 6.594049958841275e-12}, {"id": 2, "seek": 0, "start": 10.24, "end": 16.12, "text": "第一,他把全球開源模型的參數記錄直接翻了一個倍。", "tokens": [50877, 18049, 11, 5000, 16075, 11319, 28533, 8949, 47402, 41908, 39823, 1546, 47375, 30622, 16958, 9567, 226, 43297, 42716, 2289, 8990, 35477, 1543, 51171], "temperature": 0, "avg_logprob": -0.23796453147098937, "compression_ratio": 1.1478599221789882, "no_speech_prob": 6.594049958841275e-12}, {"id": 3, "seek": 0, "start": 16.12, "end": 20.22, "text": "K3模型的參數是2.8萬億。", "tokens": [51171, 42, 18, 41908, 39823, 1546, 47375, 30622, 1541, 17, 13, 23, 14099, 25459, 1543, 51376], "temperature": 0, "avg_logprob": -0.23796453147098937, "compression_ratio": 1.1478599221789882, "no_speech_prob": 6.594049958841275e-12}, {"id": 4, "seek": 0, "start": 20.48, "end": 26.88, "text": "什麼概念呢?在之前的開源陣營裡面,最大的是Deepseek,只有1.6萬億。", "tokens": [51389, 7598, 26181, 33679, 6240, 30, 3581, 32442, 1546, 8949, 47402, 8842, 96, 24184, 253, 32399, 11, 8661, 3582, 24620, 11089, 595, 405, 916, 11, 35244, 16, 13, 21, 14099, 25459, 1543, 51709], "temperature": 0, "avg_logprob": -0.23796453147098937, "compression_ratio": 1.1478599221789882, "no_speech_prob": 6.594049958841275e-12}, {"id": 5, "seek": 2688, "start": 26.88, "end": 29.7, "text": "K3的大小差不多是他的兩倍。", "tokens": [50365, 42, 18, 1546, 3582, 7322, 37876, 1541, 31309, 16313, 35477, 1543, 50506], "temperature": 0, "avg_logprob": -0.22504733541737432, "compression_ratio": 1.1764705882352942, "no_speech_prob": 9.37489235547373e-12}, {"id": 6, "seek": 2688, "start": 29.96, "end": 36.86, "text": "第二,在一項衡量真實知識工作的評測裡面,他排到了全球第三。", "tokens": [50519, 19693, 11, 3581, 2257, 8313, 227, 9890, 94, 26748, 6303, 10376, 6498, 43143, 41315, 1546, 39115, 9592, 105, 32399, 11, 5000, 44647, 21381, 11319, 28533, 35878, 1543, 50864], "temperature": 0, "avg_logprob": -0.22504733541737432, "compression_ratio": 1.1764705882352942, "no_speech_prob": 9.37489235547373e-12}, {"id": 7, "seek": 2688, "start": 37.12, "end": 42.5, "text": "第一名和第二名分別是Cloud5和GPT5.6瘦。", "tokens": [50877, 18049, 15940, 12565, 19693, 15940, 6627, 16158, 1541, 32787, 20, 12565, 38, 47, 51, 20, 13, 21, 163, 246, 99, 1543, 51146], "temperature": 0, "avg_logprob": -0.22504733541737432, "compression_ratio": 1.1764705882352942, "no_speech_prob": 9.37489235547373e-12}, {"id": 8, "seek": 2688, "start": 42.76, "end": 47.620000000000005, "text": "而且要注意的是,Cloud是OPUS4.8被他擠到了後面一名。", "tokens": [51159, 22942, 4275, 40794, 24620, 11, 32787, 1541, 12059, 3447, 19, 13, 23, 23238, 5000, 162, 6646, 21381, 5661, 8833, 2257, 15940, 1543, 51402], "temperature": 0, "avg_logprob": -0.22504733541737432, "compression_ratio": 1.1764705882352942, "no_speech_prob": 9.37489235547373e-12}, {"id": 9, "seek": 2688, "start": 48.120000000000005, "end": 52.739999999999995, "text": "開源模型第一次站到了B原刺旗艦模型的前面。", "tokens": [51427, 8949, 47402, 41908, 39823, 18049, 9487, 34155, 21381, 33, 19683, 2437, 118, 4479, 245, 12531, 99, 41908, 39823, 1546, 8945, 8833, 1543, 51658], "temperature": 0, "avg_logprob": -0.22504733541737432, "compression_ratio": 1.1764705882352942, "no_speech_prob": 9.37489235547373e-12}, {"id": 10, "seek": 5274, "start": 52.74, "end": 59.82, "text": "那麼這期視頻呢,我們就來詳細的了解一下KIMI K3,它到底強在哪?價格又有多恒?", "tokens": [50365, 20504, 2664, 16786, 27333, 29940, 6240, 11, 5884, 3111, 3763, 6716, 111, 41588, 1546, 2289, 17278, 8861, 42, 6324, 40, 591, 18, 11, 11284, 33883, 22752, 3581, 17028, 30, 45441, 30921, 17047, 2412, 6392, 11845, 240, 30, 50719], "temperature": 0, "avg_logprob": -0.23774174946110424, "compression_ratio": 1.1933085501858736, "no_speech_prob": 1.6960923984732368e-11}, {"id": 11, "seek": 5274, "start": 60.24, "end": 61.74, "text": "先來看一下時間線。", "tokens": [50740, 10108, 3763, 28324, 20788, 31699, 1543, 50815], "temperature": 0, "avg_logprob": -0.23774174946110424, "compression_ratio": 1.1933085501858736, "no_speech_prob": 1.6960923984732368e-11}, {"id": 12, "seek": 5274, "start": 62.160000000000004, "end": 68.74000000000001, "text": "其實在7月15號,KIMI的開放平台的一個促銷頁面已經提前泄露了。", "tokens": [50836, 14139, 3581, 22, 6939, 5211, 18616, 11, 42, 6324, 40, 1546, 8949, 12744, 16716, 15433, 1546, 8990, 7792, 225, 42041, 115, 8313, 223, 8833, 25338, 20949, 8945, 6847, 226, 18594, 110, 2289, 1543, 51165], "temperature": 0, "avg_logprob": -0.23774174946110424, "compression_ratio": 1.1933085501858736, "no_speech_prob": 1.6960923984732368e-11}, {"id": 13, "seek": 5274, "start": 68.74000000000001, "end": 71.56, "text": "大家就知道可能有新的模型要上了。", "tokens": [51165, 6868, 3111, 7758, 16657, 2412, 12560, 1546, 41908, 39823, 4275, 5708, 2289, 1543, 51306], "temperature": 0, "avg_logprob": -0.23774174946110424, "compression_ratio": 1.1933085501858736, "no_speech_prob": 1.6960923984732368e-11}, {"id": 14, "seek": 5274, "start": 71.94, "end": 76.42, "text": "那麼在16號,也就是昨天的晚上,KIMI K3正式上線。", "tokens": [51325, 20504, 3581, 6866, 18616, 11, 6404, 5620, 47404, 6135, 1546, 50157, 11, 42, 6324, 40, 591, 18, 15789, 27584, 5708, 31699, 1543, 51549], "temperature": 0, "avg_logprob": -0.23774174946110424, "compression_ratio": 1.1933085501858736, "no_speech_prob": 1.6960923984732368e-11}, {"id": 15, "seek": 7642, "start": 76.42, "end": 80.8, "text": "分別在KIMI的APP以及KIMI Code裡面都可以使用了。", "tokens": [50365, 6627, 16158, 3581, 42, 6324, 40, 1546, 8749, 40282, 42, 6324, 40, 15549, 32399, 7182, 6723, 22982, 9254, 2289, 1543, 50584], "temperature": 0, "avg_logprob": -0.14733367427702873, "compression_ratio": 1.277602523659306, "no_speech_prob": 1.4906301887274154e-11}, {"id": 16, "seek": 7642, "start": 81.04, "end": 87.3, "text": "現在只要進入網站APP,包括如果你是KIMI Code的會員,都可以正式使用KIMI K3了。", "tokens": [50596, 12648, 14003, 4275, 18214, 14028, 26352, 34155, 8749, 11, 41828, 13119, 32526, 42, 6324, 40, 15549, 1546, 6236, 20992, 11, 7182, 6723, 15789, 27584, 22982, 9254, 42, 6324, 40, 591, 18, 2289, 1543, 50909], "temperature": 0, "avg_logprob": -0.14733367427702873, "compression_ratio": 1.277602523659306, "no_speech_prob": 1.4906301887274154e-11}, {"id": 17, "seek": 7642, "start": 87.58, "end": 89.7, "text": "這個時間點呢,也非常的微妙。", "tokens": [50923, 6287, 20788, 8216, 6240, 11, 6404, 48263, 39152, 5648, 247, 1543, 51029], "temperature": 0, "avg_logprob": -0.14733367427702873, "compression_ratio": 1.277602523659306, "no_speech_prob": 1.4906301887274154e-11}, {"id": 18, "seek": 7642, "start": 89.96000000000001, "end": 93.76, "text": "正好卡在上海的世界人工智能大會前面。", "tokens": [51042, 15789, 2131, 32681, 3581, 5708, 20078, 1546, 24486, 4035, 23323, 5094, 118, 8225, 3582, 6236, 8945, 8833, 1543, 51232], "temperature": 0, "avg_logprob": -0.14733367427702873, "compression_ratio": 1.277602523659306, "no_speech_prob": 1.4906301887274154e-11}, {"id": 19, "seek": 7642, "start": 94.08, "end": 97.42, "text": "而且呢,月之岸面也正在談新的一輪融資。", "tokens": [51248, 22942, 6240, 11, 6939, 9574, 23182, 116, 8833, 6404, 15789, 3581, 48022, 12560, 1546, 2257, 12290, 103, 43772, 235, 35380, 1543, 51415], "temperature": 0, "avg_logprob": -0.14733367427702873, "compression_ratio": 1.277602523659306, "no_speech_prob": 1.4906301887274154e-11}, {"id": 20, "seek": 7642, "start": 97.76, "end": 100.26, "text": "估值呢,大約是315億美元。", "tokens": [51432, 7384, 108, 40242, 6240, 11, 3582, 30889, 1541, 18, 5211, 25459, 39433, 1543, 51557], "temperature": 0, "avg_logprob": -0.14733367427702873, "compression_ratio": 1.277602523659306, "no_speech_prob": 1.4906301887274154e-11}, {"id": 21, "seek": 7642, "start": 100.26, "end": 104.22, "text": "KIMI K3這個成績單是給誰看的,就不用我多說了。", "tokens": [51557, 42, 6324, 40, 591, 18, 6287, 11336, 21030, 122, 24227, 1541, 17798, 20757, 4200, 1546, 11, 3111, 24384, 1654, 6392, 4622, 2289, 1543, 51755], "temperature": 0, "avg_logprob": -0.14733367427702873, "compression_ratio": 1.277602523659306, "no_speech_prob": 1.4906301887274154e-11}, {"id": 22, "seek": 10422, "start": 104.22, "end": 106.22, "text": "然後來看一下啊,規格。", "tokens": [50365, 10213, 3763, 28324, 4905, 11, 2888, 237, 30921, 1543, 50465], "temperature": 0, "avg_logprob": -0.21455340189476535, "compression_ratio": 1.2039660056657224, "no_speech_prob": 1.8340440277597736e-11}, {"id": 23, "seek": 10422, "start": 106.22, "end": 109.82, "text": "2.8萬億這個參數呢,肯定是非常吸引眼球的。", "tokens": [50465, 17, 13, 23, 14099, 25459, 6287, 47375, 30622, 6240, 11, 14356, 107, 12088, 1541, 14392, 46431, 41301, 25281, 28533, 1546, 1543, 50645], "temperature": 0, "avg_logprob": -0.21455340189476535, "compression_ratio": 1.2039660056657224, "no_speech_prob": 1.8340440277597736e-11}, {"id": 24, "seek": 10422, "start": 109.82, "end": 115.34, "text": "但是我覺得比大更值得看的,是他怎麼樣把這麼大的模型給他跑起來。", "tokens": [50645, 11189, 17689, 11706, 3582, 19002, 40242, 5916, 4200, 1546, 11, 1541, 5000, 34253, 16075, 21269, 39156, 41908, 39823, 17798, 5000, 32585, 21670, 1543, 50921], "temperature": 0, "avg_logprob": -0.21455340189476535, "compression_ratio": 1.2039660056657224, "no_speech_prob": 1.8340440277597736e-11}, {"id": 25, "seek": 10422, "start": 115.34, "end": 120.34, "text": "看官方放的這張結構圖呢,我們也可以看到有兩個新的東西。", "tokens": [50921, 4200, 31929, 9249, 12744, 1546, 2664, 18889, 17144, 43362, 2523, 244, 6240, 11, 5884, 47935, 18032, 2412, 34623, 12560, 1546, 26978, 1543, 51171], "temperature": 0, "avg_logprob": -0.21455340189476535, "compression_ratio": 1.2039660056657224, "no_speech_prob": 1.8340440277597736e-11}, {"id": 26, "seek": 10422, "start": 120.34, "end": 125.42, "text": "一個叫KDA,KIMI Delta Tension混合線性注意力。", "tokens": [51171, 8990, 19855, 42, 7509, 11, 42, 6324, 40, 18183, 314, 3378, 48640, 14245, 31699, 21686, 40794, 13486, 1543, 51425], "temperature": 0, "avg_logprob": -0.21455340189476535, "compression_ratio": 1.2039660056657224, "no_speech_prob": 1.8340440277597736e-11}, {"id": 27, "seek": 10422, "start": 125.42, "end": 133.9, "text": "簡單解釋就是,傳統注意力的計算量呢,會隨著上下文長度平方級上漲,越長呢就越貴。", "tokens": [51425, 31995, 17278, 5873, 233, 5620, 11, 40852, 33725, 40794, 13486, 1546, 30114, 19497, 26748, 6240, 11, 6236, 48890, 19382, 5708, 4438, 17174, 15353, 13127, 16716, 9249, 35315, 5708, 14065, 110, 11, 25761, 15353, 6240, 3111, 25761, 11561, 112, 1543, 51849], "temperature": 0, "avg_logprob": -0.21455340189476535, "compression_ratio": 1.2039660056657224, "no_speech_prob": 1.8340440277597736e-11}, {"id": 28, "seek": 13390, "start": 133.9, "end": 140.06, "text": "那麼K3的做法是呢,是每四層裡面,三層用便宜的線性注意力。", "tokens": [50365, 20504, 42, 18, 1546, 10907, 11148, 1541, 6240, 11, 1541, 23664, 19425, 9636, 97, 32399, 11, 10960, 9636, 97, 9254, 27364, 2415, 250, 1546, 31699, 21686, 40794, 13486, 1543, 50673], "temperature": 0, "avg_logprob": -0.1830155508858817, "compression_ratio": 1.1967871485943775, "no_speech_prob": 2.1513922587867995e-11}, {"id": 29, "seek": 13390, "start": 140.06, "end": 142.78, "text": "然後一層用完整的全注意力。", "tokens": [50673, 10213, 2257, 9636, 97, 9254, 14128, 27662, 1546, 11319, 40794, 13486, 1543, 50809], "temperature": 0, "avg_logprob": -0.1830155508858817, "compression_ratio": 1.1967871485943775, "no_speech_prob": 2.1513922587867995e-11}, {"id": 30, "seek": 13390, "start": 142.78, "end": 147.18, "text": "這裡省下來的算力,換成一個東西,100萬Token的上下文。", "tokens": [50809, 18771, 2862, 223, 27769, 1546, 19497, 13486, 11, 36338, 11336, 8990, 26978, 11, 6879, 14099, 51, 8406, 1546, 5708, 4438, 17174, 1543, 51029], "temperature": 0, "avg_logprob": -0.1830155508858817, "compression_ratio": 1.1967871485943775, "no_speech_prob": 2.1513922587867995e-11}, {"id": 31, "seek": 13390, "start": 147.18, "end": 152.62, "text": "之前KIMI K2的上下文長度呢,只有256K,直接翻了4倍。", "tokens": [51029, 32442, 42, 6324, 40, 591, 17, 1546, 5708, 4438, 17174, 15353, 13127, 6240, 11, 35244, 6074, 21, 42, 11, 43297, 42716, 2289, 19, 35477, 1543, 51301], "temperature": 0, "avg_logprob": -0.1830155508858817, "compression_ratio": 1.1967871485943775, "no_speech_prob": 2.1513922587867995e-11}, {"id": 32, "seek": 13390, "start": 152.62, "end": 155.5, "text": "另外一個新的東西叫注意力殘差。", "tokens": [51301, 26202, 8990, 12560, 1546, 26978, 19855, 40794, 13486, 15976, 246, 21679, 1543, 51445], "temperature": 0, "avg_logprob": -0.1830155508858817, "compression_ratio": 1.1967871485943775, "no_speech_prob": 2.1513922587867995e-11}, {"id": 33, "seek": 15550, "start": 155.5, "end": 160.46, "text": "讓深層網絡把淺層的注意力信息呢,直接接力上來。", "tokens": [50365, 21195, 24043, 9636, 97, 26352, 6948, 94, 16075, 14176, 118, 9636, 97, 1546, 40794, 13486, 17665, 26460, 6240, 11, 43297, 14468, 13486, 5708, 3763, 1543, 50613], "temperature": 0, "avg_logprob": -0.1258602369399298, "compression_ratio": 1.1047120418848169, "no_speech_prob": 2.2265836743806666e-11}, {"id": 34, "seek": 15550, "start": 160.46, "end": 166.78, "text": "官方說,這套組合讓整體的訓練效率做到了KIMI K2的2.5倍。", "tokens": [50613, 31929, 9249, 4622, 11, 2664, 1881, 245, 30052, 14245, 21195, 27662, 23987, 1546, 5396, 241, 47027, 43076, 44866, 10907, 21381, 42, 6324, 40, 591, 17, 1546, 17, 13, 20, 35477, 1543, 50929], "temperature": 0, "avg_logprob": -0.1258602369399298, "compression_ratio": 1.1047120418848169, "no_speech_prob": 2.2265836743806666e-11}, {"id": 35, "seek": 15550, "start": 166.78, "end": 171.34, "text": "然後就是混合專家系統這塊,做的也非常非常的激進。", "tokens": [50929, 10213, 5620, 48640, 14245, 46072, 5155, 25368, 33725, 2664, 26268, 11, 10907, 1546, 6404, 14392, 48263, 26373, 222, 18214, 1543, 51157], "temperature": 0, "avg_logprob": -0.1258602369399298, "compression_ratio": 1.1047120418848169, "no_speech_prob": 2.2265836743806666e-11}, {"id": 36, "seek": 17134, "start": 171.34, "end": 176.78, "text": "整個模型呢,有896個專家,每次推理呢只激活16個。", "tokens": [50365, 27662, 3338, 41908, 39823, 6240, 11, 2412, 21115, 21, 3338, 46072, 5155, 11, 23664, 9487, 33597, 13876, 6240, 14003, 26373, 222, 25956, 6866, 3338, 1543, 50637], "temperature": 0, "avg_logprob": -0.11168310615453828, "compression_ratio": 1.1893203883495145, "no_speech_prob": 1.9650859064967996e-11}, {"id": 37, "seek": 17134, "start": 176.78, "end": 184.22, "text": "也就是說,雖然總參數量是2.8萬億,但是真正幹活的每次呢,只是裡面很小的一部分。", "tokens": [50637, 6404, 26222, 11, 6306, 244, 5823, 26575, 47375, 30622, 26748, 1541, 17, 13, 23, 14099, 25459, 11, 11189, 6303, 15789, 29932, 25956, 1546, 23664, 9487, 6240, 11, 36859, 32399, 4563, 7322, 1546, 2257, 32174, 1543, 51009], "temperature": 0, "avg_logprob": -0.11168310615453828, "compression_ratio": 1.1893203883495145, "no_speech_prob": 1.9650859064967996e-11}, {"id": 38, "seek": 17134, "start": 184.22, "end": 189.26, "text": "再來看一下成績單,我挑三個重點的在視頻裡面說一下。", "tokens": [51009, 35961, 28324, 11336, 21030, 122, 24227, 11, 1654, 49067, 10960, 3338, 12624, 8216, 1546, 3581, 27333, 29940, 32399, 4622, 8861, 1543, 51261], "temperature": 0, "avg_logprob": -0.11168310615453828, "compression_ratio": 1.1893203883495145, "no_speech_prob": 1.9650859064967996e-11}, {"id": 39, "seek": 18926, "start": 189.26, "end": 198.14, "text": "第一個呢,是剛才視頻開頭提到的GPTWALL,他測的呢不是做題,是44個職業裡面真實的工作任務。", "tokens": [50365, 40760, 6240, 11, 1541, 16940, 18888, 27333, 29940, 8949, 19921, 20949, 4511, 1546, 38, 47, 51, 54, 15921, 11, 5000, 9592, 105, 1546, 6240, 7296, 10907, 14475, 11, 1541, 13912, 3338, 8171, 115, 27119, 32399, 6303, 10376, 1546, 41315, 26443, 39517, 1543, 50809], "temperature": 0, "avg_logprob": -0.1905102085422825, "compression_ratio": 1.1240601503759398, "no_speech_prob": 2.2969230684122266e-11}, {"id": 40, "seek": 18926, "start": 198.14, "end": 202.94, "text": "那麼K3呢,是拿了1687分,全球第三名。", "tokens": [50809, 20504, 42, 18, 6240, 11, 1541, 24351, 2289, 6866, 23853, 6627, 11, 11319, 28533, 35878, 15940, 1543, 51049], "temperature": 0, "avg_logprob": -0.1905102085422825, "compression_ratio": 1.1240601503759398, "no_speech_prob": 2.2969230684122266e-11}, {"id": 41, "seek": 18926, "start": 202.94, "end": 214.45999999999998, "text": "第二個呢,是GP2A Diamond,研究生級別的科學題,他的得分呢是93.5%,是目前所有看源模型裡面得分最高的。", "tokens": [51049, 49085, 6240, 11, 1541, 38, 47, 17, 32, 26593, 11, 23230, 242, 44704, 8244, 35315, 16158, 1546, 42091, 21372, 14475, 11, 31309, 5916, 6627, 6240, 1541, 26372, 13, 20, 8923, 1541, 39004, 39300, 4200, 47402, 41908, 39823, 32399, 5916, 6627, 8661, 12979, 1546, 1543, 51625], "temperature": 0, "avg_logprob": -0.1905102085422825, "compression_ratio": 1.1240601503759398, "no_speech_prob": 2.2969230684122266e-11}, {"id": 42, "seek": 21446, "start": 214.46, "end": 226.46, "text": "第三個呢,我覺得是最實在的BlowComp,在網上檢索問題的答案,得分呢是91.2%,也是當時公開成績裡面最高的。", "tokens": [50365, 35878, 3338, 6240, 11, 17689, 1541, 8661, 10376, 3581, 1546, 33, 14107, 34, 8586, 11, 3581, 26352, 5708, 40741, 95, 7732, 95, 17197, 1546, 28223, 28899, 11, 5916, 6627, 6240, 1541, 29925, 13, 17, 8923, 22021, 13118, 6611, 13545, 8949, 11336, 21030, 122, 32399, 8661, 12979, 1546, 1543, 50965], "temperature": 0, "avg_logprob": -0.16359417612959698, "compression_ratio": 1.2267657992565055, "no_speech_prob": 1.818595447844462e-11}, {"id": 43, "seek": 21446, "start": 226.46, "end": 230.62, "text": "那麼這一下呢,很明顯是吃到了100萬上下文的紅利。", "tokens": [50965, 20504, 2664, 8861, 6240, 11, 4563, 11100, 46605, 1541, 10123, 21381, 6879, 14099, 5708, 4438, 17174, 1546, 33487, 23700, 1543, 51173], "temperature": 0, "avg_logprob": -0.16359417612959698, "compression_ratio": 1.2267657992565055, "no_speech_prob": 1.818595447844462e-11}, {"id": 44, "seek": 21446, "start": 230.62, "end": 233.02, "text": "當然,測評呢也只能說明一半。", "tokens": [51173, 21707, 11, 9592, 105, 39115, 6240, 6404, 14003, 8225, 4622, 11100, 2257, 30018, 1543, 51293], "temperature": 0, "avg_logprob": -0.16359417612959698, "compression_ratio": 1.2267657992565055, "no_speech_prob": 1.818595447844462e-11}, {"id": 45, "seek": 21446, "start": 233.02, "end": 236.22, "text": "另一半呢,我們還要真實上手去測試。", "tokens": [51293, 22762, 2257, 30018, 6240, 11, 5884, 7824, 4275, 6303, 10376, 5708, 11389, 6734, 9592, 105, 22099, 1543, 51453], "temperature": 0, "avg_logprob": -0.16359417612959698, "compression_ratio": 1.2267657992565055, "no_speech_prob": 1.818595447844462e-11}, {"id": 46, "seek": 21446, "start": 236.22, "end": 238.22, "text": "這裡呢,我測試了兩個案例。", "tokens": [51453, 18771, 6240, 11, 1654, 9592, 105, 22099, 2289, 34623, 28899, 17797, 1543, 51553], "temperature": 0, "avg_logprob": -0.16359417612959698, "compression_ratio": 1.2267657992565055, "no_speech_prob": 1.818595447844462e-11}, {"id": 47, "seek": 23822, "start": 238.22, "end": 244.94, "text": "第一個測試呢,是讓它根據官方的報導,以及我自己整理的文件。", "tokens": [50365, 40760, 9592, 105, 22099, 6240, 11, 1541, 21195, 11284, 31337, 36841, 31929, 9249, 1546, 36235, 11, 40282, 1654, 17645, 27662, 13876, 1546, 17174, 20485, 1543, 50701], "temperature": 0, "avg_logprob": -0.18545256555080414, "compression_ratio": 1.2428571428571429, "no_speech_prob": 1.9989988830904082e-11}, {"id": 48, "seek": 23822, "start": 244.94, "end": 254.94, "text": "讓它直接幫我生成一份Kimi K3不低於12頁的一個PPT,幻燈篇,並且直接以網頁的形式展現。", "tokens": [50701, 21195, 11284, 43297, 32187, 1654, 8244, 11336, 2257, 36266, 42, 10121, 591, 18, 1960, 41377, 19488, 4762, 8313, 223, 1546, 8990, 17755, 51, 11, 3509, 119, 24184, 230, 20878, 229, 11, 35934, 20334, 43297, 3588, 26352, 8313, 223, 1546, 30900, 27584, 43491, 9581, 1543, 51201], "temperature": 0, "avg_logprob": -0.18545256555080414, "compression_ratio": 1.2428571428571429, "no_speech_prob": 1.9989988830904082e-11}, {"id": 49, "seek": 23822, "start": 254.94, "end": 258.86, "text": "那麼這個呢,就是它一次性完成的任務。", "tokens": [51201, 20504, 6287, 6240, 11, 5620, 11284, 27505, 21686, 41509, 1546, 26443, 39517, 1543, 51397], "temperature": 0, "avg_logprob": -0.18545256555080414, "compression_ratio": 1.2428571428571429, "no_speech_prob": 1.9989988830904082e-11}, {"id": 50, "seek": 23822, "start": 258.86, "end": 265.98, "text": "可以先從第一屏看一下,Kimi K3,它用模型第一次摸到了B原模型次旗艦的門檻。", "tokens": [51397, 6723, 10108, 21068, 18049, 9636, 237, 28324, 11, 42, 10121, 591, 18, 11, 11284, 9254, 41908, 39823, 18049, 9487, 34783, 116, 21381, 33, 19683, 41908, 39823, 9487, 4479, 245, 12531, 99, 1546, 37232, 162, 33411, 1543, 51753], "temperature": 0, "avg_logprob": -0.18545256555080414, "compression_ratio": 1.2428571428571429, "no_speech_prob": 1.9989988830904082e-11}, {"id": 51, "seek": 26598, "start": 265.98, "end": 268.54, "text": "這是等於是PPT的封面。", "tokens": [50365, 16600, 10187, 19488, 1541, 17755, 51, 1546, 1530, 223, 8833, 1543, 50493], "temperature": 0, "avg_logprob": -0.14181762351129287, "compression_ratio": 1.3687943262411348, "no_speech_prob": 2.1348986814495596e-11}, {"id": 52, "seek": 26598, "start": 268.54, "end": 277.98, "text": "然後它最主要的幾個參數,它的配色,它的佈局,還有這些數據的提取,全部是Kimi K3的,自動去完成的。", "tokens": [50493, 10213, 11284, 8661, 13557, 4275, 1546, 23575, 3338, 47375, 30622, 11, 45224, 38846, 17673, 11, 45224, 1593, 230, 34703, 11, 15569, 29869, 30622, 36841, 1546, 20949, 29436, 11, 38714, 1541, 42, 10121, 591, 18, 1546, 11, 9722, 12572, 6734, 41509, 1546, 1543, 50965], "temperature": 0, "avg_logprob": -0.14181762351129287, "compression_ratio": 1.3687943262411348, "no_speech_prob": 2.1348986814495596e-11}, {"id": 53, "seek": 26598, "start": 277.98, "end": 283.26, "text": "我覺得這個版面的佈局,以及參數的提取,配色其實都做得非常好。", "tokens": [50965, 17689, 6287, 42096, 8833, 1546, 1593, 230, 34703, 11, 40282, 47375, 30622, 1546, 20949, 29436, 11, 38846, 17673, 14139, 7182, 10907, 5916, 14392, 2131, 1543, 51229], "temperature": 0, "avg_logprob": -0.14181762351129287, "compression_ratio": 1.3687943262411348, "no_speech_prob": 2.1348986814495596e-11}, {"id": 54, "seek": 26598, "start": 283.26, "end": 286.70000000000005, "text": "而且內容呢非常的詳盡,非常的豐富。", "tokens": [51229, 22942, 28472, 25750, 6240, 48263, 6716, 111, 5419, 94, 11, 48263, 19517, 238, 47564, 1543, 51401], "temperature": 0, "avg_logprob": -0.14181762351129287, "compression_ratio": 1.3687943262411348, "no_speech_prob": 2.1348986814495596e-11}, {"id": 55, "seek": 26598, "start": 286.70000000000005, "end": 292.78000000000003, "text": "還有關於它的這個架構拆解,這種圖表設計,也全部是它一次性完成的。", "tokens": [51401, 15569, 14899, 19488, 45224, 6287, 7360, 114, 43362, 6852, 228, 17278, 11, 33005, 2523, 244, 17571, 39035, 30114, 11, 6404, 38714, 1541, 11284, 27505, 21686, 41509, 1546, 1543, 51705], "temperature": 0, "avg_logprob": -0.14181762351129287, "compression_ratio": 1.3687943262411348, "no_speech_prob": 2.1348986814495596e-11}, {"id": 56, "seek": 29278, "start": 292.78, "end": 296.29999999999995, "text": "目前看起來沒有找到很明顯的錯誤。", "tokens": [50365, 39004, 4200, 21670, 6963, 25085, 4511, 4563, 11100, 46605, 1546, 13133, 3549, 97, 1543, 50541], "temperature": 0, "avg_logprob": -0.16728656075217507, "compression_ratio": 1.2112676056338028, "no_speech_prob": 2.4910433091274164e-11}, {"id": 57, "seek": 29278, "start": 296.29999999999995, "end": 299.82, "text": "基本上可以達到一個直接使用的水平了。", "tokens": [50541, 37946, 5708, 6723, 33086, 4511, 8990, 43297, 22982, 9254, 1546, 15590, 16716, 2289, 1543, 50717], "temperature": 0, "avg_logprob": -0.16728656075217507, "compression_ratio": 1.2112676056338028, "no_speech_prob": 2.4910433091274164e-11}, {"id": 58, "seek": 29278, "start": 299.82, "end": 308.61999999999995, "text": "而且是沒有創建任何Scale的前提下,只是給它發了一個簡單的報告,以及一個Planter,讓它生成PPT,就完成了。", "tokens": [50717, 22942, 1541, 6963, 5935, 113, 34157, 26443, 11906, 16806, 1220, 1546, 8945, 20949, 4438, 11, 36859, 17798, 11284, 14637, 2289, 8990, 31995, 1546, 17803, 16846, 11, 40282, 8990, 47, 8658, 391, 11, 21195, 11284, 8244, 11336, 17755, 51, 11, 3111, 41509, 2289, 1543, 51157], "temperature": 0, "avg_logprob": -0.16728656075217507, "compression_ratio": 1.2112676056338028, "no_speech_prob": 2.4910433091274164e-11}, {"id": 59, "seek": 29278, "start": 308.61999999999995, "end": 309.97999999999996, "text": "這能力我覺得還是不錯的。", "tokens": [51157, 2664, 8225, 13486, 17689, 25583, 30967, 1546, 1543, 51225], "temperature": 0, "avg_logprob": -0.16728656075217507, "compression_ratio": 1.2112676056338028, "no_speech_prob": 2.4910433091274164e-11}, {"id": 60, "seek": 29278, "start": 309.97999999999996, "end": 313.58, "text": "然後我在測試過程當中呢,發現一個比較明顯的問題。", "tokens": [51225, 10213, 40528, 9592, 105, 22099, 8816, 29649, 13118, 5975, 6240, 11, 38927, 8990, 25174, 11100, 46605, 1546, 17197, 1543, 51405], "temperature": 0, "avg_logprob": -0.16728656075217507, "compression_ratio": 1.2112676056338028, "no_speech_prob": 2.4910433091274164e-11}, {"id": 61, "seek": 31358, "start": 313.58, "end": 317.74, "text": "就是不算是現在使用的人比較多還是什麼的。", "tokens": [50365, 5620, 1960, 19497, 1541, 12648, 22982, 9254, 29979, 25174, 6392, 25583, 7598, 1546, 1543, 50573], "temperature": 0, "avg_logprob": -0.17311903912088145, "compression_ratio": 1.2592592592592593, "no_speech_prob": 2.4422915079203023e-11}, {"id": 62, "seek": 31358, "start": 317.74, "end": 325.02, "text": "就是我發送了這樣一條Planter過去,它整個完成這個任務的耗時,時間非常的長。", "tokens": [50573, 5620, 1654, 14637, 29309, 2289, 8377, 2257, 31106, 47, 8658, 391, 36910, 11, 11284, 27662, 3338, 41509, 6287, 26443, 39517, 1546, 4450, 245, 6611, 11, 20788, 48263, 15353, 1543, 50937], "temperature": 0, "avg_logprob": -0.17311903912088145, "compression_ratio": 1.2592592592592593, "no_speech_prob": 2.4422915079203023e-11}, {"id": 63, "seek": 31358, "start": 325.02, "end": 327.09999999999997, "text": "至少花費了十幾分鐘。", "tokens": [50937, 20844, 15686, 20127, 50071, 2289, 20145, 23575, 32611, 1543, 51041], "temperature": 0, "avg_logprob": -0.17311903912088145, "compression_ratio": 1.2592592592592593, "no_speech_prob": 2.4422915079203023e-11}, {"id": 64, "seek": 31358, "start": 327.09999999999997, "end": 333.26, "text": "那這個任務如果以往是用GPT或者用Cloud去完成,我覺得這時間肯定是在10分鐘以內的。", "tokens": [51041, 4184, 6287, 26443, 39517, 13119, 3588, 29510, 1541, 9254, 38, 47, 51, 31148, 9254, 32787, 6734, 41509, 11, 17689, 2664, 20788, 14356, 107, 12088, 1541, 3581, 3279, 32611, 3588, 28472, 1546, 1543, 51349], "temperature": 0, "avg_logprob": -0.17311903912088145, "compression_ratio": 1.2592592592592593, "no_speech_prob": 2.4422915079203023e-11}, {"id": 65, "seek": 31358, "start": 333.26, "end": 335.34, "text": "因為任務並不是特別的難。", "tokens": [51349, 11471, 26443, 39517, 35934, 7296, 38446, 1546, 21192, 1543, 51453], "temperature": 0, "avg_logprob": -0.17311903912088145, "compression_ratio": 1.2592592592592593, "no_speech_prob": 2.4422915079203023e-11}, {"id": 66, "seek": 31358, "start": 335.34, "end": 337.18, "text": "然後它的內容也不是很多。", "tokens": [51453, 10213, 45224, 28472, 25750, 6404, 7296, 20778, 1543, 51545], "temperature": 0, "avg_logprob": -0.17311903912088145, "compression_ratio": 1.2592592592592593, "no_speech_prob": 2.4422915079203023e-11}, {"id": 67, "seek": 33718, "start": 337.18, "end": 341.18, "text": "用KimiK3生成任務的時間是超出了我的預期的。", "tokens": [50365, 9254, 42, 10121, 42, 18, 8244, 11336, 26443, 39517, 1546, 20788, 1541, 19869, 7781, 2289, 14200, 30156, 16786, 1546, 1543, 50565], "temperature": 0, "avg_logprob": -0.16135288774967194, "compression_ratio": 1.2982456140350878, "no_speech_prob": 1.9644987026001814e-11}, {"id": 68, "seek": 33718, "start": 341.18, "end": 346.78000000000003, "text": "然後第二個任務呢,是我讓它基於我前面兩期視頻。", "tokens": [50565, 10213, 49085, 26443, 39517, 6240, 11, 42614, 21195, 11284, 26008, 19488, 1654, 8945, 8833, 16313, 16786, 27333, 29940, 1543, 50845], "temperature": 0, "avg_logprob": -0.16135288774967194, "compression_ratio": 1.2982456140350878, "no_speech_prob": 1.9644987026001814e-11}, {"id": 69, "seek": 33718, "start": 346.78000000000003, "end": 349.58, "text": "我是創建了一個體感遊戲。", "tokens": [50845, 15914, 5935, 113, 34157, 2289, 8990, 23987, 9709, 43046, 1543, 50985], "temperature": 0, "avg_logprob": -0.16135288774967194, "compression_ratio": 1.2982456140350878, "no_speech_prob": 1.9644987026001814e-11}, {"id": 70, "seek": 33718, "start": 349.58, "end": 355.02, "text": "用谷歌的模型,它可以監控我們人體的關節、部位。", "tokens": [50985, 9254, 8897, 115, 29582, 1546, 41908, 39823, 11, 11284, 6723, 5419, 96, 48707, 5884, 4035, 23987, 1546, 14899, 27694, 1231, 13470, 11160, 1543, 51257], "temperature": 0, "avg_logprob": -0.16135288774967194, "compression_ratio": 1.2982456140350878, "no_speech_prob": 1.9644987026001814e-11}, {"id": 71, "seek": 33718, "start": 355.02, "end": 357.82, "text": "首先我是做了一個體感監視器。", "tokens": [51257, 36490, 15914, 10907, 2289, 8990, 23987, 9709, 5419, 96, 27333, 34386, 1543, 51397], "temperature": 0, "avg_logprob": -0.16135288774967194, "compression_ratio": 1.2982456140350878, "no_speech_prob": 1.9644987026001814e-11}, {"id": 72, "seek": 33718, "start": 357.82, "end": 359.82, "text": "這個在上期視頻已經講過了。", "tokens": [51397, 6287, 3581, 5708, 16786, 27333, 29940, 25338, 11932, 8816, 2289, 1543, 51497], "temperature": 0, "avg_logprob": -0.16135288774967194, "compression_ratio": 1.2982456140350878, "no_speech_prob": 1.9644987026001814e-11}, {"id": 73, "seek": 33718, "start": 359.82, "end": 363.26, "text": "包括它的項目,我已經開運到KitaHub上面去了。", "tokens": [51497, 41828, 45224, 8313, 227, 11386, 11, 1654, 25338, 8949, 32772, 4511, 42, 2786, 21150, 49750, 45190, 1543, 51669], "temperature": 0, "avg_logprob": -0.16135288774967194, "compression_ratio": 1.2982456140350878, "no_speech_prob": 1.9644987026001814e-11}, {"id": 74, "seek": 36326, "start": 363.26, "end": 368.06, "text": "後面呢,我又用GPT5.64添加了幾個體感遊戲。", "tokens": [50365, 5661, 8833, 6240, 11, 1654, 17047, 9254, 38, 47, 51, 20, 13, 19395, 14176, 119, 9990, 2289, 23575, 3338, 23987, 9709, 43046, 1543, 50605], "temperature": 0, "avg_logprob": -0.1836217659107153, "compression_ratio": 1.25, "no_speech_prob": 1.925070866604095e-11}, {"id": 75, "seek": 36326, "start": 368.06, "end": 374.7, "text": "比如有體感穿墻、星球互聯、還有這個切水果的遊戲,以及雙輪對打的遊戲。", "tokens": [50605, 36757, 2412, 23987, 9709, 13740, 123, 24228, 119, 1231, 20682, 28533, 1369, 240, 43942, 1231, 15569, 6287, 23632, 15590, 9319, 1546, 43046, 11, 40282, 6306, 247, 12290, 103, 2855, 12467, 1546, 43046, 1543, 50937], "temperature": 0, "avg_logprob": -0.1836217659107153, "compression_ratio": 1.25, "no_speech_prob": 1.925070866604095e-11}, {"id": 76, "seek": 36326, "start": 374.7, "end": 377.98, "text": "這個是之前評測Fibre5以及GPT5.6生成的。", "tokens": [50937, 41427, 32442, 39115, 9592, 105, 37, 897, 265, 20, 40282, 38, 47, 51, 20, 13, 21, 8244, 11336, 1546, 1543, 51101], "temperature": 0, "avg_logprob": -0.1836217659107153, "compression_ratio": 1.25, "no_speech_prob": 1.925070866604095e-11}, {"id": 77, "seek": 36326, "start": 377.98, "end": 385.09999999999997, "text": "然後今天呢,我是想用這個KimiK3去做一個體感消消樂的遊戲。", "tokens": [51101, 10213, 12074, 6240, 11, 15914, 7093, 9254, 6287, 42, 10121, 42, 18, 6734, 10907, 8990, 23987, 9709, 28837, 28837, 34043, 1546, 43046, 1543, 51457], "temperature": 0, "avg_logprob": -0.1836217659107153, "compression_ratio": 1.25, "no_speech_prob": 1.925070866604095e-11}, {"id": 78, "seek": 36326, "start": 385.09999999999997, "end": 389.9, "text": "那其實像這套框架,在我原來的項目裡面,它已經生成了。", "tokens": [51457, 4184, 14139, 12760, 2664, 1881, 245, 21416, 228, 7360, 114, 11, 3581, 1654, 19683, 3763, 1546, 8313, 227, 11386, 32399, 11, 11284, 25338, 8244, 11336, 2289, 1543, 51697], "temperature": 0, "avg_logprob": -0.1836217659107153, "compression_ratio": 1.25, "no_speech_prob": 1.925070866604095e-11}, {"id": 79, "seek": 38990, "start": 389.9, "end": 393.41999999999996, "text": "也就是它的模型、它的底層邏輯、框架都有了。", "tokens": [50365, 6404, 5620, 45224, 41908, 39823, 1231, 45224, 22816, 9636, 97, 3023, 237, 12290, 107, 1231, 21416, 228, 7360, 114, 48121, 2289, 1543, 50541], "temperature": 0, "avg_logprob": -0.13084353381440839, "compression_ratio": 1.364864864864865, "no_speech_prob": 1.982057400151671e-11}, {"id": 80, "seek": 38990, "start": 393.41999999999996, "end": 397.02, "text": "只是讓它在這個基礎上去增加一個體感消消樂的遊戲。", "tokens": [50541, 36859, 21195, 11284, 3581, 6287, 26008, 14971, 236, 5708, 6734, 24228, 252, 9990, 8990, 23987, 9709, 28837, 28837, 34043, 1546, 43046, 1543, 50721], "temperature": 0, "avg_logprob": -0.13084353381440839, "compression_ratio": 1.364864864864865, "no_speech_prob": 1.982057400151671e-11}, {"id": 81, "seek": 38990, "start": 397.02, "end": 399.65999999999997, "text": "那麼今天用KimiK3出現了兩個問題。", "tokens": [50721, 20504, 12074, 9254, 42, 10121, 42, 18, 7781, 9581, 2289, 34623, 17197, 1543, 50853], "temperature": 0, "avg_logprob": -0.13084353381440839, "compression_ratio": 1.364864864864865, "no_speech_prob": 1.982057400151671e-11}, {"id": 82, "seek": 38990, "start": 399.65999999999997, "end": 402.38, "text": "第一個問題呢,就是剛才我提到的。", "tokens": [50853, 40760, 17197, 6240, 11, 5620, 16940, 18888, 1654, 20949, 4511, 1546, 1543, 50989], "temperature": 0, "avg_logprob": -0.13084353381440839, "compression_ratio": 1.364864864864865, "no_speech_prob": 1.982057400151671e-11}, {"id": 83, "seek": 38990, "start": 402.38, "end": 404.7, "text": "它的耗時呢,非常非常的久。", "tokens": [50989, 45224, 4450, 245, 6611, 6240, 11, 14392, 48263, 25320, 1543, 51105], "temperature": 0, "avg_logprob": -0.13084353381440839, "compression_ratio": 1.364864864864865, "no_speech_prob": 1.982057400151671e-11}, {"id": 84, "seek": 38990, "start": 404.7, "end": 408.21999999999997, "text": "運行了差不多十幾二十分鐘,都沒有完成。", "tokens": [51105, 32772, 46606, 37876, 20145, 23575, 11217, 20145, 32611, 11, 7182, 6963, 41509, 1543, 51281], "temperature": 0, "avg_logprob": -0.13084353381440839, "compression_ratio": 1.364864864864865, "no_speech_prob": 1.982057400151671e-11}, {"id": 85, "seek": 38990, "start": 408.21999999999997, "end": 414.7, "text": "然後第二個問題呢,就是我目前開通的是Kimi的第二個等級的套餐。", "tokens": [51281, 10213, 49085, 17197, 6240, 11, 5620, 1654, 39004, 8949, 19550, 24620, 42, 10121, 1546, 49085, 10187, 35315, 1546, 1881, 245, 22217, 238, 1543, 51605], "temperature": 0, "avg_logprob": -0.13084353381440839, "compression_ratio": 1.364864864864865, "no_speech_prob": 1.982057400151671e-11}, {"id": 86, "seek": 41470, "start": 414.7, "end": 422.38, "text": "也就是一個月是99塊錢。那麼是可以在KimiCode裡面去使用這個KimiK3的模型。", "tokens": [50365, 6404, 5620, 8990, 6939, 1541, 8494, 26268, 29483, 1543, 20504, 1541, 6723, 3581, 42, 10121, 34, 1429, 32399, 6734, 22982, 9254, 6287, 42, 10121, 42, 18, 1546, 41908, 39823, 1543, 50749], "temperature": 0, "avg_logprob": -0.152562567858192, "compression_ratio": 1.3191489361702127, "no_speech_prob": 2.4441960608245772e-11}, {"id": 87, "seek": 41470, "start": 422.38, "end": 424.46, "text": "如果你開通的是第一個等級的套餐。", "tokens": [50749, 45669, 8949, 19550, 24620, 40760, 10187, 35315, 1546, 1881, 245, 22217, 238, 1543, 50853], "temperature": 0, "avg_logprob": -0.152562567858192, "compression_ratio": 1.3191489361702127, "no_speech_prob": 2.4441960608245772e-11}, {"id": 88, "seek": 41470, "start": 424.46, "end": 428.7, "text": "目前在KimiCode裡面是使用不了這個K3的模型的。", "tokens": [50853, 39004, 3581, 42, 10121, 34, 1429, 32399, 1541, 22982, 9254, 47225, 6287, 42, 18, 1546, 41908, 39823, 1546, 1543, 51065], "temperature": 0, "avg_logprob": -0.152562567858192, "compression_ratio": 1.3191489361702127, "no_speech_prob": 2.4441960608245772e-11}, {"id": 89, "seek": 41470, "start": 428.7, "end": 434.7, "text": "完事呢,我在測試第二個任務的時候,可以看到,它給我羅列的toDo裡面有這麼多步驟。", "tokens": [51065, 14128, 6973, 6240, 11, 40528, 9592, 105, 22099, 49085, 26443, 39517, 20643, 11, 6723, 18032, 11, 11284, 17798, 1654, 36004, 43338, 1546, 1353, 7653, 32399, 2412, 21269, 6392, 31429, 24023, 253, 1543, 51365], "temperature": 0, "avg_logprob": -0.152562567858192, "compression_ratio": 1.3191489361702127, "no_speech_prob": 2.4441960608245772e-11}, {"id": 90, "seek": 41470, "start": 434.7, "end": 437.98, "text": "我覺得這個步驟是設計的有點過於冷餘的。", "tokens": [51365, 17689, 6287, 31429, 24023, 253, 1541, 39035, 30114, 1546, 42440, 8816, 19488, 32499, 22217, 246, 1546, 1543, 51529], "temperature": 0, "avg_logprob": -0.152562567858192, "compression_ratio": 1.3191489361702127, "no_speech_prob": 2.4441960608245772e-11}, {"id": 91, "seek": 43798, "start": 437.98, "end": 445.98, "text": "然後在做第二個任務的時候,僅僅只是完成了第一條toDo,它5個小時的用量限制就已經達到了。", "tokens": [50365, 10213, 3581, 10907, 49085, 26443, 39517, 20643, 11, 9502, 227, 9502, 227, 36859, 41509, 2289, 18049, 31106, 1353, 7653, 11, 11284, 20, 3338, 46997, 1546, 9254, 26748, 43446, 25491, 3111, 25338, 33086, 21381, 1543, 50765], "temperature": 0, "avg_logprob": -0.1284869539326635, "compression_ratio": 1.368421052631579, "no_speech_prob": 2.7067414282155866e-11}, {"id": 92, "seek": 43798, "start": 445.98, "end": 449.66, "text": "然後並且呢,週用量也達到了20%的使用。", "tokens": [50765, 10213, 35934, 20334, 6240, 11, 38003, 9254, 26748, 6404, 33086, 21381, 2009, 4, 1546, 22982, 9254, 1543, 50949], "temperature": 0, "avg_logprob": -0.1284869539326635, "compression_ratio": 1.368421052631579, "no_speech_prob": 2.7067414282155866e-11}, {"id": 93, "seek": 43798, "start": 449.66, "end": 455.1, "text": "也就是說,我開通99塊錢一個月的會員,僅僅只是完成了一個多的任務。", "tokens": [50949, 6404, 5620, 4622, 11, 1654, 8949, 19550, 8494, 26268, 29483, 8990, 6939, 1546, 6236, 20992, 11, 9502, 227, 9502, 227, 36859, 41509, 2289, 8990, 6392, 1546, 26443, 39517, 1543, 51221], "temperature": 0, "avg_logprob": -0.1284869539326635, "compression_ratio": 1.368421052631579, "no_speech_prob": 2.7067414282155866e-11}, {"id": 94, "seek": 43798, "start": 455.1, "end": 456.46000000000004, "text": "第二個任務還沒有完成。", "tokens": [51221, 49085, 26443, 39517, 7824, 6963, 41509, 1543, 51289], "temperature": 0, "avg_logprob": -0.1284869539326635, "compression_ratio": 1.368421052631579, "no_speech_prob": 2.7067414282155866e-11}, {"id": 95, "seek": 43798, "start": 456.46000000000004, "end": 460.06, "text": "有可能是因為這個模型太大,然後現在用的人太多。", "tokens": [51289, 2412, 16657, 1541, 11471, 6287, 41908, 39823, 9455, 3582, 11, 10213, 12648, 9254, 29979, 9455, 6392, 1543, 51469], "temperature": 0, "avg_logprob": -0.1284869539326635, "compression_ratio": 1.368421052631579, "no_speech_prob": 2.7067414282155866e-11}, {"id": 96, "seek": 46006, "start": 460.06, "end": 463.06, "text": "所以它整體的一個使用量限制比較少。", "tokens": [50365, 7239, 11284, 27662, 23987, 1546, 8990, 22982, 9254, 26748, 43446, 25491, 25174, 15686, 1543, 50515], "temperature": 0, "avg_logprob": -0.16584007981894672, "compression_ratio": 1.3302752293577982, "no_speech_prob": 2.2195714016737256e-11}, {"id": 97, "seek": 46006, "start": 463.06, "end": 469.06, "text": "包括我開通的這個第二等級的會員,它並不支持100萬的上下文。", "tokens": [50515, 41828, 1654, 8949, 19550, 1546, 6287, 19693, 10187, 35315, 1546, 6236, 20992, 11, 11284, 35934, 1960, 35488, 6879, 14099, 1546, 5708, 4438, 17174, 1543, 50815], "temperature": 0, "avg_logprob": -0.16584007981894672, "compression_ratio": 1.3302752293577982, "no_speech_prob": 2.2195714016737256e-11}, {"id": 98, "seek": 46006, "start": 469.06, "end": 477.06, "text": "只有256K,那麼你必須是開通這個200塊錢一個月的,才可以享受100萬的上下文長度。", "tokens": [50815, 35244, 6074, 21, 42, 11, 20504, 2166, 28531, 8313, 230, 1541, 8949, 19550, 6287, 7629, 26268, 29483, 8990, 6939, 1546, 11, 18888, 6723, 27654, 23151, 6879, 14099, 1546, 5708, 4438, 17174, 15353, 13127, 1543, 51215], "temperature": 0, "avg_logprob": -0.16584007981894672, "compression_ratio": 1.3302752293577982, "no_speech_prob": 2.2195714016737256e-11}, {"id": 99, "seek": 46006, "start": 477.06, "end": 480.74, "text": "給我的感覺還是不知道因為算力,還是什麼其他原因啊。", "tokens": [51215, 17798, 14200, 31742, 25583, 17572, 11471, 19497, 13486, 11, 25583, 7598, 9572, 5000, 42577, 4905, 1543, 51399], "temperature": 0, "avg_logprob": -0.16584007981894672, "compression_ratio": 1.3302752293577982, "no_speech_prob": 2.2195714016737256e-11}, {"id": 100, "seek": 46006, "start": 480.74, "end": 482.74, "text": "誠意不是特別的夠。", "tokens": [51399, 3549, 254, 9042, 7296, 38446, 1546, 31649, 1543, 51499], "temperature": 0, "avg_logprob": -0.16584007981894672, "compression_ratio": 1.3302752293577982, "no_speech_prob": 2.2195714016737256e-11}, {"id": 101, "seek": 46006, "start": 482.74, "end": 484.74, "text": "接下來呢,說一下價格。", "tokens": [51499, 40012, 6240, 11, 4622, 8861, 45441, 30921, 1543, 51599], "temperature": 0, "avg_logprob": -0.16584007981894672, "compression_ratio": 1.3302752293577982, "no_speech_prob": 2.2195714016737256e-11}, {"id": 102, "seek": 46006, "start": 484.74, "end": 488.74, "text": "這個可能是對使用者、開發者殺傷力最大的部分。", "tokens": [51599, 6287, 16657, 1541, 2855, 22982, 9254, 12444, 1231, 8949, 14637, 12444, 45528, 41323, 13486, 8661, 39156, 32174, 1543, 51799], "temperature": 0, "avg_logprob": -0.16584007981894672, "compression_ratio": 1.3302752293577982, "no_speech_prob": 2.2195714016737256e-11}, {"id": 103, "seek": 48874, "start": 488.74, "end": 493.22, "text": "輸入呢,是3美元每100萬Token,輸出呢是15美元。", "tokens": [50365, 12290, 116, 14028, 6240, 11, 1541, 18, 39433, 23664, 6879, 14099, 51, 8406, 11, 12290, 116, 7781, 6240, 1541, 5211, 39433, 1543, 50589], "temperature": 0, "avg_logprob": -0.16707810052007221, "compression_ratio": 1.2582781456953642, "no_speech_prob": 2.6308597661506283e-11}, {"id": 104, "seek": 48874, "start": 493.22, "end": 496.3, "text": "如果是緩存命中的話,輸入只要3毛。", "tokens": [50589, 13119, 1541, 12608, 102, 39781, 25236, 5975, 21358, 11, 12290, 116, 14028, 14003, 4275, 18, 39057, 1543, 50743], "temperature": 0, "avg_logprob": -0.16707810052007221, "compression_ratio": 1.2582781456953642, "no_speech_prob": 2.6308597661506283e-11}, {"id": 105, "seek": 48874, "start": 496.3, "end": 499.22, "text": "這一塊呢,一直是KIMI做的比較好的地方。", "tokens": [50743, 32260, 26268, 6240, 11, 34448, 1541, 42, 6324, 40, 10907, 1546, 25174, 20715, 30146, 1543, 50889], "temperature": 0, "avg_logprob": -0.16707810052007221, "compression_ratio": 1.2582781456953642, "no_speech_prob": 2.6308597661506283e-11}, {"id": 106, "seek": 48874, "start": 499.22, "end": 502.3, "text": "OK,那麼強的地方呢,已經說完了。", "tokens": [50889, 9443, 11, 20504, 22752, 1546, 30146, 6240, 11, 25338, 4622, 41665, 1543, 51043], "temperature": 0, "avg_logprob": -0.16707810052007221, "compression_ratio": 1.2582781456953642, "no_speech_prob": 2.6308597661506283e-11}, {"id": 107, "seek": 48874, "start": 502.3, "end": 504.74, "text": "接下來呢,來潑兩盆冷水。", "tokens": [51043, 40012, 6240, 11, 3763, 35622, 239, 16313, 5419, 228, 32499, 15590, 1543, 51165], "temperature": 0, "avg_logprob": -0.16707810052007221, "compression_ratio": 1.2582781456953642, "no_speech_prob": 2.6308597661506283e-11}, {"id": 108, "seek": 48874, "start": 504.74, "end": 510.54, "text": "第一盤呢,全球最大開用模型的這個說法呢,它現在還只是一個期貨。", "tokens": [51165, 18049, 5419, 97, 6240, 11, 11319, 28533, 8661, 3582, 8949, 9254, 41908, 39823, 1546, 6287, 4622, 11148, 6240, 11, 11284, 12648, 7824, 36859, 8990, 16786, 11561, 101, 1543, 51455], "temperature": 0, "avg_logprob": -0.16707810052007221, "compression_ratio": 1.2582781456953642, "no_speech_prob": 2.6308597661506283e-11}, {"id": 109, "seek": 48874, "start": 510.54, "end": 513.94, "text": "模型能用是真的,但是呢,權重還沒有開放。", "tokens": [51455, 41908, 39823, 8225, 9254, 1541, 8034, 11, 11189, 6240, 11, 40212, 12624, 7824, 6963, 8949, 12744, 1543, 51625], "temperature": 0, "avg_logprob": -0.16707810052007221, "compression_ratio": 1.2582781456953642, "no_speech_prob": 2.6308597661506283e-11}, {"id": 110, "seek": 51394, "start": 513.94, "end": 519.4200000000001, "text": "官方是承諾在7月27號之前,放出完整的權重和技術報告。", "tokens": [50365, 31929, 9249, 1541, 3416, 123, 11067, 122, 3581, 22, 6939, 10076, 18616, 32442, 11, 12744, 7781, 14128, 27662, 1546, 40212, 12624, 12565, 32502, 9890, 241, 17803, 16846, 1543, 50639], "temperature": 0, "avg_logprob": -0.15501728887143343, "compression_ratio": 1.1285140562248996, "no_speech_prob": 2.3937836088339104e-11}, {"id": 111, "seek": 51394, "start": 519.4200000000001, "end": 524.4200000000001, "text": "第二盤呢,就算是權重開放了,我們自己本地電腦呢,也是跑不動的。", "tokens": [50639, 19693, 5419, 97, 6240, 11, 3111, 19497, 1541, 40212, 12624, 8949, 12744, 2289, 11, 5884, 17645, 8802, 10928, 20545, 21184, 99, 6240, 11, 22021, 32585, 1960, 12572, 1546, 1543, 50889], "temperature": 0, "avg_logprob": -0.15501728887143343, "compression_ratio": 1.1285140562248996, "no_speech_prob": 2.3937836088339104e-11}, {"id": 112, "seek": 51394, "start": 524.4200000000001, "end": 531.22, "text": "2.8萬億參數呢,就算你有1.5TB內存的MGDODO集群,跑起來都夠欠。", "tokens": [50889, 17, 13, 23, 14099, 25459, 47375, 30622, 6240, 11, 3111, 19497, 43320, 16, 13, 20, 51, 33, 28472, 39781, 1546, 44, 38, 35, 14632, 46, 26020, 7248, 97, 11, 32585, 21670, 7182, 31649, 5988, 254, 1543, 51229], "temperature": 0, "avg_logprob": -0.15501728887143343, "compression_ratio": 1.1285140562248996, "no_speech_prob": 2.3937836088339104e-11}, {"id": 113, "seek": 51394, "start": 531.22, "end": 533.62, "text": "這是一個數據中心級別的模型。", "tokens": [51229, 16600, 8990, 30622, 36841, 5975, 7945, 35315, 16158, 1546, 41908, 39823, 1543, 51349], "temperature": 0, "avg_logprob": -0.15501728887143343, "compression_ratio": 1.1285140562248996, "no_speech_prob": 2.3937836088339104e-11}, {"id": 114, "seek": 53362, "start": 533.62, "end": 541.5, "text": "所以,開源對於個人的意義呢,更多是讓第三方平台能便宜的把它跑起來,而不是說把它下載到本地自己去跑。", "tokens": [50365, 7239, 11, 8949, 47402, 2855, 19488, 40104, 1546, 9042, 37430, 6240, 11, 19002, 6392, 1541, 21195, 35878, 9249, 16716, 15433, 8225, 27364, 2415, 250, 1546, 42061, 32585, 21670, 11, 11070, 7296, 4622, 42061, 4438, 12290, 231, 4511, 8802, 10928, 17645, 6734, 32585, 1543, 50759], "temperature": 0, "avg_logprob": -0.16138280992922577, "compression_ratio": 1.036764705882353, "no_speech_prob": 1.970489223179772e-11}], "language": "zh"}