Mistral Large 4評測:歐洲最大模型,開源權重卻跳票 | Mistral Large 4 Review: Big Model, Delayed Open Weights
By Kit 小克 | AI Tool Observer | 2026-10-13
🇹🇼 Mistral Large 4評測:歐洲最大模型,開源權重卻跳票
Mistral Large 4(內部代號「Le Chonk」)是法國 Mistral AI 在 2026 年 10 月 6 日公開預覽的新旗艦模型,號稱是歐洲目前最大的語言模型,總參數達 1.05 兆、啟用專家層約 490~520 億參數,支援 100 萬 token 超長上下文與 160+ 語言。聽起來很猛,但實際評測顯示,Mistral Large 4 距離真正的「世界第一」還有一段距離。
規格與定價:先用再開源
目前只開放 API 預覽,官方承諾的開源權重要等到 10 月底才會釋出——也就是說,現在能用的 Mistral Large 4 其實跟「開源模型」還沾不上邊。API 定價方面,上線前兩週有 50% 折扣:輸入每百萬 token 0.68 美元、輸出 2.09 美元;折扣結束後恢復成 1.36 / 4.18 美元,跟主流一線模型相比不算便宜,也不算貴。
跑分實測:歐洲最強,但不是世界最強
獨立測試機構的結果很老實:在程式碼(Deep 1.1 拿 62 分)和代理任務(Automation Bench 59.9 分)上,Mistral Large 4 在開源模型裡排第二,輸給 DeepSeek 和 GLM 這些中國陣營的開源模型。唯一拿下第一的是資安評測(Artificial Analysis Cyber Index),和 GLM 5.3 Flash 並列冠軍。換句話說,這是歐洲目前做得出來最好的模型,但放到全球開源賽局裡,中國模型仍然領先。
誰該關注這顆模型
- 需要資料主權(歐盟法規、不想資料出境)的企業——這是 Mistral 的核心賣點
- 做資安相關應用的團隊——跑分唯一拿第一的項目
- 想要純開源權重自架的人——先等 10 月底,現在拿到的只是 API
如果你只是想要跑分最高、最便宜的模型,現階段 Mistral Large 4 還不是答案;但如果你在意「歐洲製造、未來可自架」,這是目前最接近的選項。好不好用,試了才知道。
🇺🇸 Mistral Large 4 Review: Big Model, Delayed Open Weights
Mistral Large 4, nicknamed "Le Chonk," is the flagship model French AI lab Mistral released into public preview on October 6, 2026, billed as Europe's largest language model to date: 1.05 trillion total parameters, roughly 49-52 billion active per expert, a 1-million-token context window, and native support for 160+ languages. It's an impressive spec sheet — but the honest read is that Mistral Large 4 still isn't the global leader its scale would suggest.
Pricing and Availability: API Now, Open Weights Later
Right now you can only access Mistral Large 4 through the preview API — the open weights Mistral promised aren't due until end of October, so calling this an "open-weight model" today is premature. Launch pricing carries a 50% discount for the first two weeks: $0.68 per million input tokens and $2.09 per million output tokens, reverting to $1.36 / $4.18 afterward. That puts it in the same ballpark as other frontier-tier APIs — not a bargain, not a ripoff.
Benchmarks: Best in Europe, Not Best in the World
Independent testing paints a realistic picture. On coding (Deep 1.1, score 62) and agentic tasks (Automation Bench, 59.9), Mistral Large 4 ranks second among open-weight models, behind Chinese competitors like DeepSeek and GLM. The one category where it genuinely leads is cybersecurity, tying for first on the Artificial Analysis Cyber Index alongside GLM 5.3 Flash. In short: it's the best model Europe has shipped, but the global open-model race is still led from China.
Who Should Actually Care
- Enterprises with data-sovereignty requirements (EU regulation, no data leaving the region) — this is Mistral's core pitch
- Security-focused teams — the one benchmark category where it tops the chart
- Anyone planning to self-host open weights — wait until October 31; what's available now is API-only
If you're chasing the cheapest, highest-scoring model on the market, Mistral Large 4 isn't it yet. But if "built in Europe, eventually self-hostable" matters to you, this is the closest thing available right now. You won't know until you try it.
Sources / 資料來源
- Mistral AI 官方公告:Introducing Mistral Large 4
- Tom's Hardware 獨立測試:Large 4 仍輸中國開源模型一截
- OpenRouter:Mistral Large 4 API 定價與跑分
延伸閱讀 / Related Articles
- Muse Wonder評測:Meta首款前沿模型,最強版鎖訂閱 | Muse Wonder Review: Meta's Frontier Model, Locked Tier
- QGIS評測:開源地圖軟體封殺AI程式碼貢獻 | QGIS Review: Open-Source GIS Bans AI Code Contributions
- Copilot多模型評測:微軟砸20億美元導入Claude | Copilot Multi-Model Review: Microsoft Pays Anthropic $2B
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