Kimi K3評測：2.8兆參數最大開源模型，你真的跑得動嗎 | Kimi K3 Review: Biggest Open-Weight Model You Can't Run

By Kit 小克 | AI Tool Observer | 2026-07-30

🇹🇼 Kimi K3評測：2.8兆參數最大開源模型，你真的跑得動嗎

Kimi K3 是中國新創 Moonshot AI 在 2026 年 7 月 16 日發布、7 月 27 日釋出完整權重的最新旗艦模型，號稱是史上參數量最大的開源權重模型——2.8 兆參數，超越 DeepSeek V4 Pro 的 1.6 兆，一舉刷新紀錄。但「開源」兩個字背後，藏著一個大部分人沒注意到的現實問題。

Kimi K3 是什麼：2.8 兆參數的巨獸

Kimi K3 採用混合專家（MoE）架構，896 個專家中每次僅啟動 16 個，搭配 100 萬 token 的超長上下文窗口，主打長流程程式設計、知識工作與多步驟 agent 任務。原始權重檔案高達約 1.4 TB。

跑分實測：在前端評測贏過 Claude Fable 5

Kimi K3 的成績單相當亮眼：

在 GDPval-AA v2（涵蓋 44 種職業、9 大產業的真實任務評測）拿下 1,687 分，排名第三，僅次於 Claude Fable 5 Max（1,815）與 GPT-5.6 Sol Max（1,747.8），贏過 Claude Opus 4.8（1,600）。
在 Arena 的 Frontend Code（前端程式碼）盲測中，Kimi K3 以 1,679 分暫居第一，超越 Claude Fable 5。
知名開發者 Simon Willison 用經典「畫一隻騎腳踏車的鵜鶘」測試，K3 順利產出有效 SVG，花費 16,658 個輸出 token、成本約 0.25 美元。

「開源」的陷阱：一般人根本跑不動

這是 Kimi K3 最值得誠實討論的地方。Moonshot AI 建議的正式部署配置是 64 張以上加速卡組成的超節點，單張消費級或單張資料中心 GPU 都塞不下這顆模型。換句話說，雖然權重公開下載，實際上能自架的只有雲端業者與大型企業，一般開發者或中小團隊碰不到「自己跑」這個選項。

API 價格：跟 Claude Sonnet 差不多，不算便宜

多數人只能透過 API 使用 Kimi K3：快取命中輸入每百萬 token 0.30 美元，未命中則是 3 美元，輸出每百萬 token 15 美元——定價幾乎與 Claude Sonnet 打平，是中國模型陣營裡最貴的一個，比自家上一代 K2.6 貴了約兩倍。目前也只提供單一「max」推理強度，導致實際使用時 token 消耗偏高。

Kit 小克怎麼看

Kimi K3 證明了中國團隊在算力受限下，仍能用巨大 MoE 架構逼近甚至局部超越 GPT-5.6 與 Claude Fable 5 這類前沿模型，前端程式碼能力尤其突出。但「開源權重」在這個規模下更像是象徵意義大於實用意義——真正能自架的只有雲端巨頭，多數團隊還是得乖乖付 API 費用，而且價格並不比閉源模型便宜。如果你只是想找一顆便宜好用的開源模型，K3 可能不是首選；但如果你需要頂級前端程式碼生成能力，又不想被單一供應商綁死，它值得放進評估清單。

好不好用，試了才知道。

🇺🇸 Kimi K3 Review: Biggest Open-Weight Model You Can't Run

Kimi K3, released by Chinese startup Moonshot AI on July 16, 2026 (full weights published July 27), is now the largest open-weight model ever shipped at 2.8 trillion parameters — surpassing DeepSeek V4 Pro's 1.6 trillion. But the word "open" hides a practical catch most coverage glosses over.

What Kimi K3 Actually Is

Kimi K3 uses a mixture-of-experts (MoE) architecture, activating just 16 of 896 experts per inference pass, paired with a 1-million-token context window aimed at long-horizon coding, knowledge work, and multi-step agent tasks. The raw weight files run to roughly 1.4 TB.

Benchmarks: Beating Claude Fable 5 on Frontend Code

The results are genuinely strong:

On GDPval-AA v2, a real-world task benchmark spanning 44 occupations and 9 industries, Kimi K3 scored 1,687 — third place, behind only Claude Fable 5 Max (1,815) and GPT-5.6 Sol Max (1,747.8), and ahead of Claude Opus 4.8 (1,600).
On Arena's blind Frontend Code evaluation, K3 currently ranks first at 1,679, ahead of Claude Fable 5.
Developer Simon Willison ran his classic "SVG pelican riding a bicycle" test — K3 produced a valid result using 16,658 output tokens for about /bin/zsh.25.

The Open-Weight Catch: Most People Still Can't Run It

This is the part worth being honest about. Moonshot recommends a supernode of 64 or more accelerators for production deployment — the model simply doesn't fit on a single consumer or even single datacenter GPU. So while the weights are technically downloadable, self-hosting is realistically only an option for cloud providers and large enterprises. For most developers and small teams, "open-weight" doesn't mean "runs on my hardware."

API Pricing: About the Same as Claude Sonnet

Most users will access Kimi K3 through the API instead: /bin/zsh.30 per million cached input tokens, per million on cache misses, and per million output tokens — pricing that lands close to Claude Sonnet and makes K3 the most expensive model to come out of a Chinese lab, roughly double its predecessor K2.6. It also currently ships with only one reasoning-effort level ("max"), which drives up real-world token consumption.

Kit's Take

Kimi K3 proves Chinese labs can still push frontier-class results under compute constraints by scaling MoE architecture, and its frontend coding performance is genuinely impressive. But at this scale, "open weights" is more symbolic than practical — real self-hosting is reserved for cloud giants, while everyone else pays API prices that aren't actually cheaper than closed models. If you want an affordable open model to run yourself, K3 probably isn't it. But if you need top-tier frontend code generation without vendor lock-in, it's worth adding to your evaluation list.

好不好用，試了才知道。

Sources / 資料來源

延伸閱讀 / Related Articles

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

Claude Code 實測：AI 幫你寫程式到底行不行？ | Claude Code Review: Can AI Really Code for You?

3月 26, 2026

閱讀完整內容

搜尋此網誌

AI小貼士