Kimi K3開源模型評測:2.8兆參數寫程式贏過Claude | Kimi K3 Review: China's 2.8T Model Beats Claude at Coding
By Kit 小克 | AI Tool Observer | 2026-08-02
🇹🇼 Kimi K3開源模型評測:2.8兆參數寫程式贏過Claude
Kimi K3是中國Moonshot AI在2026年7月底推出的開源大型語言模型,參數量高達2.8兆,是目前公開釋出過最大的開源模型,在多項寫程式與代理型(agentic)任務測試中打敗Claude Fable 5與GPT-5.6 Sol,成為這幾天開發者圈討論度最高的AI話題。
Kimi K3是什麼?跟前一代K2差在哪?
Kimi K3是Moonshot AI最新旗艦模型,第一段直接回答:它採用稀疏混合專家(MoE)架構,2.8兆參數中每次只啟動一小部分,兼顧效能與運算成本。權重已於7月27日上架Hugging Face,共96個分片、約1.56TB,任何人都能下載、微調、甚至商用部署。
Kimi K3寫程式真的贏Claude嗎?
直接回答:是的,在獨立的Frontend Code Arena排行榜上,Kimi K3拿下1,679 Elo分數排名第一,超越Claude Fable 5(1,631)與GPT-5.6 Sol(1,618)。以下是幾項關鍵成績:
- Frontend Code Arena:1,679 Elo,排名第一
- GPQA Diamond:93.5%,開源模型史上最高分
- 自動化任務(Automation Bench、SpreadsheetBench 2、BrowseComp):8項中拿下4項第一
- GDPval-AA v2 綜合評測:全球第三,僅落後Fable 5 Max與GPT-5.6 Sol Max
Kimi K3的API收費怎麼算?
直接回答:每百萬輸入token 3美元、輸出15美元,命中快取降到0.3美元。這個價格橫跨完整104萬token的超長上下文視窗,沒有分級加價,比多數旗艦模型的收費簡單許多。
開源授權有什麼限制?
直接回答:Kimi K3並非採用MIT等寬鬆授權,而是自訂的「Kimi K3 License」——可自由使用、修改、散佈、商用,但年營收超過2,000萬美元的模型即服務業者需另簽授權,月活躍用戶破1億的產品則要在畫面上掛名Kimi K3的貢獻。
小克實測心得
身為常態評測AI工具的觀察者,Kimi K3最吸引人的不是跑分數字,而是「開源+便宜+效能夠強」三者兼得。如果你已經在用開源模型跑agentic pipeline,Kimi K3值得列入候選清單;但2.8兆參數代表自架成本極高,多數團隊還是得透過API或雲端服務使用,這跟真正「本地可跑」的小模型完全是兩回事,別被「開源」兩個字誤導成免費本地部署。
好不好用,試了才知道。
🇺🇸 Kimi K3 Review: China's 2.8T Model Beats Claude at Coding
Kimi K3 is Moonshot AI's new 2.8-trillion-parameter open-weight language model, released in China in late July 2026. It's the largest open-source model ever shipped, and it's already beating Claude Fable 5 and GPT-5.6 Sol on several coding and agentic benchmarks — which is why it's the AI story developers are actually talking about this week.
What Is Kimi K3, and How Is It Different from K2?
Short answer: Kimi K3 is Moonshot AI's new flagship model, built on a sparse mixture-of-experts (MoE) architecture that activates only a fraction of its 2.8 trillion parameters per token. The weights landed on Hugging Face on July 27 — 96 shards, roughly 1.56TB — open for anyone to download, fine-tune, or deploy commercially.
Does Kimi K3 Actually Beat Claude at Coding?
Short answer: yes. On the independent Frontend Code Arena leaderboard, Kimi K3 tops the ranking with 1,679 Elo, ahead of Claude Fable 5 (1,631) and GPT-5.6 Sol (1,618).
- Frontend Code Arena: 1,679 Elo, ranked #1
- GPQA Diamond: 93.5%, the highest score ever published by an open-weight model
- Automation Bench, SpreadsheetBench 2, BrowseComp: first place in 4 of 8 real-world task benchmarks
- GDPval-AA v2: third overall, behind only Fable 5 Max and GPT-5.6 Sol Max
How Much Does the Kimi K3 API Cost?
Short answer: $3 per million input tokens and $15 per million output tokens, dropping to $0.30 on cache hits — flat pricing across the full 1,048,576-token context window, no length-based tiers.
What Are the Licensing Restrictions?
Short answer: Kimi K3 doesn't ship under MIT — it uses a custom Kimi K3 License. You can use, modify, distribute, and sell commercially, but model-as-a-service businesses earning over $20M in any 12-month period need a separate agreement, and products with more than 100 million monthly users must credit Kimi K3 on-screen.
Kit's Take
The interesting part isn't the benchmark scores — it's that Kimi K3 pulls off open weights, low API pricing, and genuinely strong performance at the same time. If you're already running agentic pipelines on open models, Kimi K3 is worth a spot on your shortlist. But 2.8 trillion parameters means self-hosting is out of reach for most teams — you'll be using it through an API or cloud provider, not running it locally like a small open model. Don't let "open source" fool you into thinking it's free to self-host.
好不好用,試了才知道。 Try it yourself — that's the only way to know.
Sources / 資料來源
- VentureBeat: China's Moonshot AI releases Kimi K3, the largest open-source model ever
- Tom's Hardware: Moonshot's 2.8-trillion-parameter Kimi K3 beats Claude Fable 5 in Frontend Code Arena
- OpenRouter: Kimi K3 API Pricing & Benchmarks
常見問題 FAQ
Kimi K3是免費的嗎?可以本地部署嗎?
權重公開下載免費,但2.8兆參數的硬體需求極高,一般團隊難以本地自架,實務上多半還是透過官方API或雲端服務商使用。
Kimi K3跟Kimi K2差在哪裡?
K3延續稀疏混合專家(MoE)架構並大幅擴增參數量與訓練資料,在寫程式、代理型任務與知識類基準測試上全面優於K2。
Kimi K3可以商用嗎?
可以,但需遵守自訂的Kimi K3 License:年營收超過2,000萬美元的模型即服務業者需另簽授權,月活躍用戶破1億的產品要掛名致謝。
Kimi K3真的比Claude、GPT厲害嗎?
在特定基準測試(如Frontend Code Arena寫程式排行榜、GPQA Diamond)確實拿下第一,但在GDPval-AA v2等綜合任務評測上仍落後Fable 5 Max與GPT-5.6 Sol Max,屬於「特定領域最強」而非全面超越。
延伸閱讀 / Related Articles
- Claude Sonnet 5漲價倒數:帳單恐暴增50% | Claude Sonnet 5 Price Deadline: Bills May Jump 50%
- Chrome AI抓漏洞實測:1,072個漏洞一次修復、13年老洞現形 | Chrome AI Fixes 1,072 Bugs, Finds 13-Year-Old Flaw
- OpenAI Astra模型評測:多代理協作解開10道數學懸案 | OpenAI Astra: New Model Solved 10 Unsolved Math Problems
AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends
留言
張貼留言