跳到主要內容

Mistral Large 4評測:開源權重放出,授權卻留一手 | Mistral Large 4 Review: Open Weights, Not Open Source

By Kit 小克 | AI Tool Observer | 2026-10-30

🇹🇼 Mistral Large 4評測:開源權重放出,授權卻留一手

Mistral Large 4(暱稱「Le Chonk」)在10月底正式放出開源權重,法國 Mistral AI 一出手就是 1 兆參數的怪物模型,號稱是全球最強的開源程式碼模型,還打敗了中國的 GLM-5.3。消息一上 Hacker News 就衝破 900 多分,討論度很高。但認真看完細節,這次的「開源」兩個字,建議打個問號。

Mistral Large 4 是什麼?

簡單講,Mistral Large 4 是一個總參數 1.05 兆、實際啟用 52B 的混合專家模型,主打程式碼生成、資安分析跟長任務推理。官方數據顯示它在 DeepSWE v1.1 拿下 62%,程式碼能力略勝 GLM-5.3,而且只用了對手 47% 的 token 量就打平甚至反超,等於單次任務算力成本省一大截。

真的是「開源」嗎?

這是整件事最值得注意的地方:Mistral Large 4 的權重放在 Hugging Face,用的是「修改版 MIT 授權」,個人或企業內部使用免費,但如果你想拿它架設服務對外營利(as a service),就得另外跟 Mistral 談商業授權。Hacker News 上一堆人在吵這算不算真的「open source」,比較準確的說法是「開源權重(open-weight)」——你能下載、能檢視、能微調,但不是完全沒有限制的開放原始碼。上一代 Large 3 用的是標準 Apache 2.0,這次換了授權條款,等於自由度實際上是退步的。

一般開發者自己架得動嗎?

答案很現實:架不動。1 兆參數即使是量化版,官方建議至少要 4 張 Nvidia B200 或 B300 才跑得順,這等級的硬體一般工作室、個人開發者根本碰不到。想用的話,務實做法是透過 Mistral 自家 API(La Plateforme)或 Hugging Face 的雲端推理,而不是妄想在自己電腦上本地跑。

  • 總參數 1.05 兆,啟用參數 52B,含 1.6B 視覺編碼器
  • DeepSWE v1.1 程式碼測試 62%,略勝 GLM-5.3 且省 47% token
  • 授權:修改版 MIT,內部商用免費,對外服務需另購授權
  • 自架門檻:量化版仍需 4 張 B200/B300 等級 GPU

結論是:Mistral Large 4 在程式碼能力跟效率上確實有實力,是歐洲陣營難得打進第一梯隊的開源模型,但「開源」這個詞被用得有點寬鬆,商用授權跟硬體門檻都不是個人開發者能輕鬆跨過的門。好不好用,試了才知道。


🇺🇸 Mistral Large 4 Review: Open Weights, Not Open Source

Mistral Large 4 (nicknamed "Le Chonk") dropped its open weights in late October, and France's Mistral AI didn't go small — a 1-trillion-parameter model claiming to be the world's best open-weight coding model, edging out China's GLM-5.3. The release shot past 900 points on Hacker News within hours. But once you read the fine print, the word "open" here deserves a second look.

What Is Mistral Large 4?

Mistral Large 4 is a mixture-of-experts model with 1.05 trillion total parameters and 52B active, built for code generation, security analysis, and long-context reasoning. Mistral reports 62% on DeepSWE v1.1, putting it slightly ahead of GLM-5.3 — while using roughly 47% fewer tokens per task, a real cost advantage for anyone running it at scale.

Is It Actually Open Source?

This is the part worth pausing on. The weights sit on Hugging Face under a modified MIT license: free for personal use and internal commercial deployment, but running it as a hosted service for others requires a separate commercial license from Mistral. Hacker News commenters were quick to point out this makes it "open-weight," not "open-source" in the strict sense — you can download, inspect, and fine-tune it, but the freedoms aren't unconditional. Large 3 shipped under plain Apache 2.0; Large 4's new terms are a step back in openness, not forward.

Can You Actually Self-Host It?

Realistically, no. Even the quantized version needs at least four Nvidia B200 or B300 GPUs to run comfortably — far beyond what any solo developer or small team has sitting around. The practical path is Mistral's own API (La Plateforme) or Hugging Face-hosted inference, not a dream of running this on a home rig.

  • 1.05T total parameters, 52B active, plus a 1.6B vision encoder
  • 62% on DeepSWE v1.1, beating GLM-5.3 while using 47% fewer tokens
  • License: modified MIT — free internally, paid license required to offer it as a service
  • Self-hosting floor: 4x B200/B300-class GPUs even for the quantized build

Mistral Large 4 is a legitimate technical win for Europe's open-model camp, genuinely strong on coding and efficiency. But "open" is being stretched here — the licensing and hardware bar both put it out of reach for most individual developers. You won't know until you try it.

Sources / 資料來源

常見問題 FAQ

Mistral Large 4 是真的開源嗎?

不完全是。權重用「修改版 MIT 授權」放在 Hugging Face,內部使用免費,但若要拿來對外提供服務則需另購商業授權,嚴格說是「開源權重」而非完整開源。

一般開發者能在自己電腦跑 Mistral Large 4 嗎?

幾乎不可能。總參數達 1.05 兆,即使是量化版,官方建議也要 4 張 Nvidia B200 或 B300 等級 GPU 才跑得動,遠超個人或小團隊的硬體能力。

Mistral Large 4 跟 GLM-5.3 比起來誰更強?

在 DeepSWE v1.1 程式碼測試上 Mistral Large 4 拿下 62%,略勝 GLM-5.3,且只用對手約 47% 的 token 量,效率明顯更好。

商用 Mistral Large 4 要額外付費嗎?

內部商用(自己公司用)免費,但如果要把它包裝成服務對外銷售(as a service),就需要另外向 Mistral 取得商業授權。

要用 Mistral Large 4 該去哪裡?

權重公開在 Hugging Face 可供下載研究,一般使用建議透過 Mistral 自家 API 平台 La Plateforme 呼叫,不建議自行本地部署。

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

這個網誌中的熱門文章

Google Ironwood TPU v7 推理專用晶片解析:效能追平 NVIDIA、成本低 44%,AI 晶片戰爭正式開打 | Google Ironwood TPU v7 Explained: Matching NVIDIA Performance at 44% Lower Cost — The AI Chip War Heats Up

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

Cursor vs GitHub Copilot vs Claude Code:AI 程式助手大比拼 | AI Coding Assistants Compared: Cursor vs GitHub Copilot vs Claude Code