Kimi K3 評測：2.8兆參數開源模型贏過 Claude 與 GPT | Kimi K3 Review: China's 2.8T Open Model Tops Claude, GPT
By Kit 小克 | AI Tool Observer | 2026-07-26
🇹🇼 Kimi K3 評測：2.8兆參數開源模型贏過 Claude 與 GPT
Moonshot AI 於 2026 年 7 月正式推出 Kimi K3，一款高達 2.8 兆參數的開源模型，是目前史上最大的開放權重模型。完整模型權重預計在 7 月 27 日全面公開，在多項指標上直接對打 Claude 與 GPT 等閉源旗艦模型，甚至在部分測試中反超，讓 AI 圈瞬間炸鍋。
Kimi K3 是什麼？
Kimi K3 是中國新創 Moonshot AI 打造的混合專家（MoE）架構開源模型，總參數量 2.8 兆，但每次推論只啟動 896 個專家中的 16 個，約 1.8%，藉此壓低運算成本。模型支援 100 萬 token 的超長上下文視窗，主打長文本推理與 AI 代理（agent）工作流。
兩項架構創新：KDA 與 AttnRes
Kimi K3 導入 Kimi Delta Attention（KDA）與 Attention Residuals（AttnRes）兩項注意力機制改良，官方表示這能同時提升推論效率與長鏈推理品質，是這次能以更低啟用成本逼近旗艦模型表現的關鍵。
Kimi K3 效能如何？真的贏過 Claude 和 GPT 嗎？
在 Frontend Code Arena 前端程式碼競技榜上，Kimi K3 以 1679 分拿下第一，超過 Claude Fable 5（1631 分）、GPT-5.6 Sol（1618 分）與 GLM-5.2（1587 分）。這是開源模型首次在該榜單登頂，證明 Kimi K3 的程式碼生成能力已站上第一線。
Kimi K3 怎麼用？API 定價多少？
Kimi K3 的 API 定價為每百萬輸入 token 3 美元、每百萬輸出 token 15 美元，是中國各家 AI 實驗室中定價最高的一款，但換算每個任務的實際成本，仍比 Anthropic Opus 4.8 便宜約一半。完整權重採 MXFP4 量化格式釋出，檔案體積約 1.4TB，開發者可自行下載部署，或透過官方 API 直接呼叫。
值得關注的重點
- 開源規模：2.8 兆參數，史上最大開放權重模型
- 推論效率：MoE 架構僅啟動約 1.8% 專家，兼顧規模與成本
- 長文本能力：支援 100 萬 token 上下文視窗
- 實測成績：Frontend Code Arena 登頂，超越 Claude 與 GPT
對開發者來說，Kimi K3 的意義不只是「又一個大模型」，而是開源陣營第一次在關鍵榜單上真正壓過閉源旗艦。但大型 MoE 模型的部署門檻仍然不低，中小團隊多半還是得透過 API 使用，實際體驗仍有待更多開發者驗證。好不好用，試了才知道。
常見問題 FAQ
Q: Kimi K3 是完全免費的嗎？
A: 權重開源可自行下載部署，但透過官方 API 使用仍需付費，定價為每百萬輸入 3 美元、輸出 15 美元。
Q: Kimi K3 和 Kimi K2.7 Code 有什麼不同？
A: K2.7 Code 是先前針對編碼任務優化的版本，K3 則是全新一代 2.8 兆參數旗艦模型，架構與規模都大幅提升。
Q: 一般開發者能自己部署 Kimi K3 嗎？
A: 技術上可行，但 1.4TB 的權重檔案需要相當硬體資源，多數中小團隊建議直接透過 API 使用。
🇺🇸 Kimi K3 Review: China's 2.8T Open Model Tops Claude, GPT
China's Moonshot AI officially launched Kimi K3 in July 2026, a 2.8-trillion-parameter open-weight model — now the largest open-weight release in history. Full weights are set to release on July 27, and the model already goes toe-to-toe with closed flagship systems from Anthropic and OpenAI, beating them outright on at least one major benchmark.
What Is Kimi K3?
Kimi K3 is a mixture-of-experts (MoE) model built by Chinese AI lab Moonshot AI, totaling 2.8 trillion parameters but activating only 16 of 896 experts per token — roughly 1.8% of the full model. It ships with a 1-million-token context window, aimed squarely at long-horizon reasoning and agentic coding workloads.
Two Architecture Upgrades: KDA and AttnRes
The model introduces two attention mechanism improvements — Kimi Delta Attention (KDA) and Attention Residuals (AttnRes) — which Moonshot says boost both inference efficiency and long-chain reasoning quality, letting the sparse model approach flagship-level performance without activating anywhere near its full parameter count.
Does Kimi K3 Actually Beat Claude and GPT?
On the Frontend Code Arena leaderboard, Kimi K3 scored 1,679 points, topping Claude Fable 5 (1,631), GPT-5.6 Sol (1,618), and GLM-5.2 (1,587). It's the first time an open-weight model has taken the number one spot on that board, a concrete signal that open-source coding capability has caught up to the frontier.
How Do You Use Kimi K3, and What Does It Cost?
API pricing is set at $3 per million input tokens and $15 per million output tokens — the highest of any Chinese lab's pricing, but still roughly half the per-task cost of Anthropic's Opus 4.8. The full weights ship in MXFP4 quantization at around 1.4 terabytes, so developers can self-host or call the model through Moonshot's API.
Key Takeaways
- Scale: 2.8 trillion parameters — the largest open-weight model ever released
- Efficiency: MoE architecture activates only ~1.8% of experts per token
- Context: 1-million-token context window for long documents and agents
- Benchmark: #1 on Frontend Code Arena, ahead of both Claude and GPT
Kimi K3 matters less as "another big model" and more as the first time an open-weight system has genuinely topped a serious leaderboard over closed frontier models. Self-hosting a 1.4TB model is still out of reach for most small teams, so API access will likely be how most people actually try it. 好不好用，試了才知道 — you won't know until you use it yourself.
FAQ
Q: Is Kimi K3 free to use?
A: The weights are open and free to self-host, but Moonshot's official API is paid — $3 per million input tokens and $15 per million output tokens.
Q: How is Kimi K3 different from Kimi K2.7 Code?
A: K2.7 Code was a coding-focused variant of the prior generation. K3 is an entirely new 2.8-trillion-parameter flagship with a larger architecture and expanded scale.
Q: Can individual developers realistically self-host Kimi K3?
A: Technically yes, but the 1.4TB weight file demands serious hardware. Most small teams will use the API instead.
Sources / 資料來源
- VentureBeat: China's Moonshot AI releases Kimi K3, the largest open-source model ever
- Tom's Hardware: Moonshot releases 2.8-trillion-parameter Kimi K3
- Interconnects.ai: Kimi K3 — The open-weights escalation
常見問題 FAQ
Kimi K3 是完全免費的嗎？
權重開源可自行下載部署，但透過官方 API 使用仍需付費，定價為每百萬輸入 3 美元、輸出 15 美元。
Kimi K3 和 Kimi K2.7 Code 有什麼不同？
K2.7 Code 是先前針對編碼任務優化的版本，K3 則是全新一代 2.8 兆參數旗艦模型，架構與規模都大幅提升。
一般開發者能自己部署 Kimi K3 嗎？
技術上可行，但 1.4TB 的權重檔案需要相當硬體資源，多數中小團隊建議直接透過 API 使用。
延伸閱讀 / Related Articles
- DeepSeek V4 正式上線：百萬字元視窗標配，尖峰時段變貴兩倍 | DeepSeek V4 Goes GA: 1M Context Standard, Peak-Hour Pricing Doubles Cost
- CISA用Mythos AI稽核政府軟體：五角大廈才封殺Anthropic | CISA Uses Anthropic's Mythos to Audit Government Code
- 黃仁勳首發X挺開源AI：50企業聯署反封殺令 | Jensen Huang's X Debut Backs Open-Weight AI
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