Mistral 開放萬億參數旗艦 Large 4 權重,改採修改版 MIT 授權,僅 MaaS 轉售須另談授權Mistral opens the weights of its trillion-parameter Large 4 flagship under a modified MIT licence, requiring a separate deal only for hosted resale
- 事件
- 法國 Mistral 於十月二十九日開放 Mistral Large 4 權重。模型總參數一萬億、每 token 激活四百九十億,原生多模態,用歐洲自有數據中心三千八百塊 Grace Blackwell 訓練。France's Mistral released the weights of Mistral Large 4 on 29 October. The natively multimodal model carries one trillion total parameters with 49 billion active per token, trained on 3,800 Grace Blackwell GPUs in its own European data centres.
- 背後
- 模型自十月六日預覽後追加三週強化學習,主攻編碼與防禦性網安。開源陣營要證明前沿能力不必綁定單一供應商。Three further weeks of reinforcement learning since the 6 October preview targeted coding and defensive cybersecurity. Open weights must show frontier capability need not be tied to one vendor, which resonates most in sectors that must demonstrate control.
我們認為開放權重的真正賣點是不會被切斷的存取權,而非跑分。企業應把它當作談判籌碼:保留一條自託管路線,即使短期不用,也能壓低閉源方案的漲價空間。We see the real selling point of open weights as access that cannot be cut off, not leaderboard position. Treat it as negotiating leverage: keep a self-hosted route on file and it caps the pricing power of closed vendors even if unused.
- 對日常工作的影響
- 資訊主管要把運行門檻算清:FP8 要八張 B300、壓縮版要四張 GB300,硬件與運維成本須與 API 報價同表比較,否則自託管的帳面優勢會在第二年消失。IT leads must price the floor: FP8 needs eight B300 GPUs and the compressed build four GB300s. Put hardware and operations in the same table as API rates or the apparent saving disappears in year two.
- 如何改善
- 本週挑一個已經上線、對外可斷的工序,試算自託管與 API 兩條路線的年成本,寫成半頁備忘給財務。This week, take one live workflow that can be cut off externally, cost the self-hosted and API routes over a year, and send finance a half-page memo.
