跳到主要內容

Kimi K3評測：2.8兆參數開源模型免費挑戰Fable 5 | Kimi K3 Review: 2.8T Open-Weight Model Rivals Claude Fable 5

By Kit 小克 | AI Tool Observer | 2026-08-08

🇹🇼 Kimi K3評測：2.8兆參數開源模型免費挑戰Fable 5

Kimi K3 是中國新創 Moonshot AI 在 2026 年 7 月推出的開源大模型，總參數量達 2.8 兆，號稱是全球最大的開放權重 AI 模型，重點是它完全免費下載、任何人都能自己部署，直接挑戰 Anthropic 和 OpenAI 的閉源旗艦模型。這篇文章帶你看 Kimi K3 到底強在哪、跑得動嗎、值不值得換。

Kimi K3 是什麼？跟其他開源模型差在哪？

Kimi K3 是一個原生多模態的混合專家（MoE）模型，896 個專家中每次推理只啟動 16 個，支援高達 100 萬 token 的超長上下文，能同時處理文字、圖片、影片。Moonshot AI 自稱它是「全球首個開放的 3T 級系統」，這規模比 Meta Llama 系列和阿里 Qwen 系列的旗艦版都要大上一截。

跑分表現如何？真的打得贏一線模型嗎？

根據 Moonshot AI 公布的評測，Kimi K3 在程式碼與代理（agentic）任務上贏過 Claude Opus 4.8 和 GPT 5.5，甚至在 Frontend Code Arena 這類前端程式碼跑分中一度超越 Claude Fable 5。不過官方也坦承，整體綜合能力仍落後 Anthropic 的 Claude Fable 5 和 OpenAI 的 GPT 5.6 Sol，尤其在需要長時間規劃的複雜推理任務上還有差距。換句話說，它不是全面碾壓，而是在特定領域（尤其是寫程式、長文件分析、跑很久的自動化任務）表現亮眼。

Kimi K3 自架成本高嗎？一般人跑得起來嗎？

自己部署 Kimi K3 門檻不低：需要至少 8 張以上的 GB300 或 MI350X/MI355X 等級 GPU，模型權重就佔了約 1.56TB，官方建議用 64 張加速卡起跳的叢集規模來跑正式環境。這不是一般開發者筆電或單卡工作站能碰的等級，比較適合有雲端資源的團隊。

  • 自架硬體門檻：8+ 張旗艦級 GPU，權重約 1.56TB
  • API 價格：每百萬輸入 token 3 美元、輸出 15 美元，快取命中可降到 0.3 美元
  • 推理框架：官方已釋出 vLLM 上的 Kimi Delta Attention 支援，方便串接

如果沒有自架需求，直接走 API 反而更划算，價格跟一線模型比起來相對親民，適合想省成本又要長上下文能力的團隊。

常見問題 FAQ

整體來說，Kimi K3 證明了開源模型也能在特定任務上逼近甚至超越閉源旗艦，尤其中國團隊在算力受限下仍能做出 3T 級模型，是這波開源浪潮很值得關注的指標。但自架門檻高、綜合能力仍非頂級，適合誰用要看你的實際需求。

好不好用，試了才知道。

🇺🇸 Kimi K3 Review: 2.8T Open-Weight Model Rivals Claude Fable 5

Kimi K3 is a 2.8-trillion-parameter open-weight model released by Chinese startup Moonshot AI in July 2026 — the largest openly available AI model to date, and it's free to download and self-host, putting direct pressure on closed frontier models from Anthropic and OpenAI. Here's what it actually gets right, what it costs to run, and whether it's worth your attention.

What Is Kimi K3 and How Is It Different from Other Open Models?

Kimi K3 is a natively multimodal Mixture-of-Experts model that activates just 16 of its 896 experts per inference pass, supports a 1-million-token context window, and handles text, images, and video natively. Moonshot AI calls it "the world's first open 3T-class system" — meaningfully larger than flagship open releases from Meta's Llama line or Alibaba's Qwen series.

Does It Actually Beat Top-Tier Models?

According to Moonshot's own benchmarks, Kimi K3 outperforms Claude Opus 4.8 and GPT 5.5 on coding and agentic tasks, and even edged out Claude Fable 5 on the Frontend Code Arena benchmark. That said, Moonshot itself admits K3 still trails Anthropic's Claude Fable 5 and OpenAI's GPT 5.6 Sol on overall performance, particularly on complex reasoning tasks that require long-horizon planning. In short: not an across-the-board win, but genuinely strong at coding, long-document analysis, and multi-hour agentic automation.

How Much Does It Cost to Run Kimi K3 Yourself?

Self-hosting isn't for hobbyists. You'll need at least 8 flagship-class GPUs (GB300 or MI350X/MI355X tier), model weights alone run around 1.56TB, and Moonshot recommends supernode clusters of 64+ accelerators for production. This is squarely infrastructure-team territory, not a single-GPU workstation project.

  • Self-hosting bar: 8+ flagship GPUs, ~1.56TB of weights
  • API pricing: $3 per million input tokens, $15 per million output, dropping to $0.30 on cache hits
  • Inference support: Production-ready Kimi Delta Attention support now shipped for vLLM

For most teams without dedicated GPU infrastructure, the API is the practical entry point — priced competitively against frontier closed models while offering that huge context window.

FAQ

Kimi K3 shows open-weight models can now match or beat closed frontier models on specific tasks like coding and long-context agentic work — a notable milestone given China's compute constraints. It's not a universal upgrade over closed leaders, and self-hosting demands serious infrastructure, so whether it's right for you depends on your workload.

好不好用，試了才知道。 (You won't know if it's good until you've actually tried it.)

Sources / 資料來源

常見問題 FAQ

Kimi K3 是免費的嗎？

模型權重完全免費開放下載，可自行部署；若不想自架,也可透過 API 付費使用，價格為每百萬輸入 token 3 美元、輸出 15 美元。

Kimi K3 能打贏 Claude Fable 5 嗎？

在 Frontend Code Arena 等特定程式碼跑分曾超越，但 Moonshot AI 自己承認整體綜合能力仍落後 Claude Fable 5 和 GPT 5.6 Sol。

一般人可以在自己電腦跑 Kimi K3 嗎？

不行。官方建議至少 8 張旗艦級 GPU（GB300 或 MI350X/MI355X），模型權重約 1.56TB，屬於企業級部署規模。

Kimi K3 的上下文長度有多長？

支援高達 100 萬 token 的上下文窗口，適合處理長文件分析與長時間運行的自動化任務。

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

留言

張貼留言