Mistral Large 4 open weights land: 1T MoE you can self-host

The 1T-param multimodal MoE is downloadable under a near-MIT license, priced $0.70/$2.10 per 1M on the API, and ~47% more token-efficient than rivals.

Nowline NOV 4 1:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Weights on Hugging Face — serve it with vLLM or SGLang

    Mistral Large 4 "Le Chonk" is a ~1T-param (52B active) natively multimodal MoE with a 1M-token context, now downloadable from Hugging Face. Self-host it with vLLM or SGLang, or grab the NVFP4 quant and community GGUF builds to run a Western frontier model with no API dependency.

  • $0.70 in / $2.10 out — and ~47% fewer tokens per task

    On the Mistral API (mistral-large-4) it's $0.70 per 1M input tokens, $0.07 cached, and $2.10 output. Mistral says it ties GLM-5.3 on the Artificial Analysis Coding Agent Index (48%) while spending ~47% fewer tokens per task — so your real coding-agent bill drops below the sticker price.

  • Top Western open-weight on the intelligence index

    It claims the highest Artificial Analysis Intelligence Index v5 score (26) of any Western open-weight model, 41% on Terminal-Bench 4.0 (up 13 points from the October preview), and 51% on the Cyber Index. These are Mistral's own numbers — benchmark it on your own stack before you re-pin.

  • License catch: use it freely, don't resell it as a service

    The Mistral Ecosystem Open-Weight License is a modified MIT — free to fine-tune and ship inside commercial products — with one clause that bites: running a Model-as-a-Service on it needs a separate commercial license from Mistral. Self-host inside your own product, yes; resell raw inference, no.