跳到主要內容

Mistral Large 4評測:1兆參數開源模型,權重10/27才給 | Mistral Large 4 Review: 1T Open Model, Weights Land Oct 27

By Kit 小克 | AI Tool Observer | 2026-10-07

🇹🇼 Mistral Large 4評測:1兆參數開源模型,權重10/27才給

Mistral Large 4是法國AI新創Mistral在2026年10月6日發布的旗艦模型,總參數高達1.05兆(其中真正啟用的只有490億),號稱是歐洲目前最大的「開源」模型,要挑戰GPT-6.1 Sol、Gemini 4 Argon和中國的GLM、DeepSeek系列。但重點是:現在能用的只有API,開源權重官方說要等到10月底才會放出來。

Mistral Large 4是什麼?規格一次看

Mistral Large 4(內部代號「Le Chonk」)是一個原生多模態的混合專家(MoE)模型,用大約4000張Nvidia Grace Blackwell GPU、花兩個月在歐洲自家機房訓練完成。

  • 總參數:1.05兆,啟用參數約490億
  • 視覺編碼器:16億參數,支援圖片輸入
  • 官方標稱context window:100萬token
  • API定價:每百萬輸入token約0.68美元起、輸出2.09美元起,比GPT-6.1 Sol和Gemini 4 Argon的2美元/10美元門檻便宜不少

Mistral Large 4效能打得過GPT-6.1、Gemini 4嗎?

官方公布的跑分顯示Mistral Large 4在部分項目有競爭力,但沒有全面贏過對手。

  • 程式能力:DeepSWE v1.1拿61.7%,但Terminal-Bench 4只有28.3%
  • 金融測試FinWorkBench:67%,跟DeepSeek-V4-Pro同級,小贏GLM-5.3的65%
  • 資安測試Cybench:93%,CyberGym-E2E(抓漏洞後還要補丁):82%
  • 視覺定位(在圖片裡找物件):DIOR-RSVG測試73%,比GPT-6 Astra的68%略高

號稱「開源權重」,為什麼現在下載不到?

這是這次發布最尷尬的地方:Mistral對外強調Large 4是「開源承諾」的延續,但官網的權重欄位現在寫著「Coming soon」。Mistral自己說會在月底放出來,VentureBeat和Euronews拿到的具體日期是10月27日。也就是說,從宣布到真正能自架、能微調,中間足足空窗三週以上,這段時間只能透過Mistral的API或Ollama雲端標籤使用,不是真正能下載到本機的模型。

現在值得用Mistral Large 4嗎?小克的實測建議

如果你在等一個能自己架在機房裡的歐洲開源大模型,Mistral Large 4現在還做不到,先別急著規劃基礎設施。但如果只是要用API寫程式或做金融、資安相關任務,目前的折扣價確實比GPT-6.1 Sol和Gemini 4 Argon划算,值得先用免費額度測測看跟自己工作流合不合。等10月27日權重真的放出來,再決定要不要花資源自架。

好不好用,試了才知道。


🇺🇸 Mistral Large 4 Review: 1T Open Model, Weights Land Oct 27

Mistral Large 4 is the new flagship model from French AI startup Mistral, announced on October 6, 2026. It packs 1.05 trillion total parameters (only 49B active per token) and is pitched as Europe's biggest "open-weight" answer to GPT-6.1 Sol, Gemini 4 Argon, and China's GLM/DeepSeek lineup. The catch: right now you can only reach it through Mistral's API — the actual open weights aren't out yet.

What Is Mistral Large 4, Exactly?

Mistral Large 4 (internally nicknamed "Le Chonk") is a natively multimodal mixture-of-experts model trained over two months on roughly 4,000 Nvidia Grace Blackwell GPUs inside Mistral's own European data centers.

  • Total parameters: 1.05T, with ~49B active per token
  • Vision encoder: 1.6B parameters for image input
  • Advertised context window: 1 million tokens
  • API pricing: from $0.68/M input and $2.09/M output tokens — undercutting the $2/$10 floor set by GPT-6.1 Sol and Gemini 4 Argon

How Does Mistral Large 4 Perform Against GPT-6.1 and Gemini 4?

Mistral's own benchmarks show Mistral Large 4 is competitive in a few areas, but it doesn't sweep the board.

  • Coding: 61.7% on DeepSWE v1.1, but just 28.3% on Terminal-Bench 4
  • Finance: 67% on FinWorkBench, tying DeepSeek-V4-Pro and edging past GLM-5.3's 65%
  • Security: 93% on Cybench, 82% on the harder CyberGym-E2E patch-and-reproduce test
  • Vision grounding: 73% on DIOR-RSVG, slightly ahead of GPT-6 Astra's 68%

If It's "Open-Weight," Why Can't I Download It Yet?

This is the awkward part: Mistral markets Large 4 as proof it's still committed to open weights, but the download page currently just says "Coming soon." Mistral says weights ship by the end of October, and reports point to a specific date of October 27. That leaves a three-week-plus gap between the announcement and anything you can actually self-host or fine-tune — for now it's API-only, or a cloud-hosted tag on Ollama, not a real local download.

Is Mistral Large 4 Worth Using Right Now?

If you're waiting for a genuinely self-hostable European open model, Mistral Large 4 isn't there yet — hold off on infrastructure planning. But if you just need an API for coding, finance, or security-adjacent tasks, the discounted pricing already undercuts GPT-6.1 Sol and Gemini 4 Argon, so it's worth testing against your workflow with the free tier now. Save the self-hosting decision for after October 27, when the weights actually land.

好不好用,試了才知道。
Good or not, you won't know until you try it.

Sources / 資料來源

常見問題 FAQ

Mistral Large 4的開源權重什麼時候會公開?

Mistral官方表示10月底前會釋出,外媒報導的具體日期是2026年10月27日,現在只能透過API或Ollama雲端標籤使用,還不能下載到本機。

Mistral Large 4跟GPT-6.1 Sol、Gemini 4 Argon比誰強?

各有勝負:Mistral Large 4在資安測試Cybench拿93%、金融測試贏過GLM-5.3,但程式能力的Terminal-Bench 4只有28.3%,沒有全面領先對手。

Mistral Large 4的API價格多少?

促銷價每百萬輸入token約0.68美元、輸出2.09美元起,比GPT-6.1 Sol和Gemini 4 Argon的2美元/10美元門檻便宜不少。

Mistral Large 4的模型規模有多大?

總參數1.05兆,但每次推論只啟用約490億參數,屬於混合專家(MoE)架構,另外還搭了16億參數的視覺編碼器。

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