Mistral released the weights of its 1T-parameter multimodal MoE on October 29. It leads open-weight coding and cyber benchmarks, runs on 4× GB300 via NVFP4, and uses a modified MIT license restricting only Model-as-a-Service resale.

The Drop: October 29, 2026

Mistral AI published the open weights of Mistral Large 4 (nicknamed Le Chonk) on Hugging Face on October 29, three weeks after an API-only public preview ^[S1]. This is the first European-trained trillion-parameter model whose weights you can download, inspect, fine-tune, and deploy on your own infrastructure ^[S1]^[S2].

We release the weights of Le Chonk, advancing the frontier of open models in agentic coding and cybersecurity.

The release includes two checkpoints:

FP8 (block-scaled weights) — the reference precision
NVFP4 (NVFP4 experts, FP8 attention/shared experts) — targets Blackwell GPUs, 606 GB mixed-precision checkpoint matching FP8 on evaluations ^[S3]^[S4]

Both are available at mistralai/Mistral-Large-4-1T-A52B and mistralai/Mistral-Large-4-1T-A52B-NVFP4 on Hugging Face ^[S2].

—

What You’re Getting: Architecture at a Glance

The model uses DeepSeek-style sparse attention (MLA + DSA top-k indexer) with several novel pieces: cross-layer KV sharing (9 NoPE layers read layer 35’s latent cache), NoPE layers (every 4th layer, zero RoPE dims), SSMax query scaling, and LeRoPE learned per-frequency scaling ^[S3]^[S4]. Day-0 support landed in vLLM (model runner V2, breakable CUDA graphs) and SGLang ^[S4].

—

Benchmarks: What the Numbers Say (Self-Reported)

All figures below are from Mistral’s technical report and blog; independent reproduction is pending ^[S1]^[S3].

Coding Agents

Cybersecurity

General & Reasoning

BrowseComp (web browsing): 84.5% (vs 75% at SFT checkpoint) — higher than GLM-5.3 or DeepSeek Flash 4.1 1017 ^[S1]
Decision Index 0.4: Not directly reported; Drex Max v1 (different model) scores 67.52
Multilingual: Native fluency in 160+ languages, all EU official languages ^[S2]

Caveat: These are vendor-reported scores on selected benchmarks. The technical report acknowledges that similar language-modeling loss can coexist with substantially different agentic results — SWE-bench Verified worsened when moving from full attention to sliding-window attention in ablation ^[S3]. Treat as directional until independent labs reproduce.

—

The License: Modified MIT with a MaaS Carve-Out

The Mistral Ecosystem Open-Weight License is a modified MIT license ^[S5]. Key terms:

Free to use, copy, modify, merge, publish, distribute, sublicense, sell copies
Free to run, deploy, further train, create derivatives
Carve-out: If you or an affiliate operate a Model-as-a-Service business (giving third parties meaningful control over inputs, parameters, or training data via API/hosted endpoints), you need a separate commercial license from Mistral
Exempt: Internal use; use via Mistral’s official products (Studio, API, Vibe Code)
No warranty, no liability — standard MIT disclaimer

This means:

✅ Self-host for your own products, internal tools, fine-tuning, on-prem deployment
✅ Embed in end-user features where the model isn’t exposed as a raw inference service
❌ Sell raw API access to Mistral Large 4 without a commercial agreement
❌ Host a “Mistral Large 4 as a Service” platform for third parties

Mistral frames this as protecting the sovereign AI category — European-trained, European-controlled, available for European enterprises to own and customize ^[S1].

—

Deployment Reality: Hardware You’ll Need

vLLM commands (from day-0 blog) ^[S4]:

“`bash

FP8 on 8 GPUs

docker pull vllm/vllm-openai:mistral-large-4 vllm serve mistralai/Mistral-Large-4-1T-A52B \ –tensor-parallel-size 8 \ –speculative-config ‘{“method”: “dspark”, “num_speculative_tokens”: 3}’ \ –tool-call-parser mistral –enable-auto-tool-choice \ –reasoning-parser mistral

NVFP4 on 4 GPUs (Blackwell)

vllm serve mistralai/Mistral-Large-4-1T-A52B-NVFP4 \ –tensor-parallel-size 4 \ –speculative-config ‘{“method”: “dspark”, “num_speculative_tokens”: 3}’ \ –tool-call-parser mistral –enable-auto-tool-choice \ –reasoning-parser mistral “`

DSpark speculative decoding (built-in 3-layer sliding-window MLA drafter) accepts ~2.9 tokens/step, ~2× output throughput on TP8 ^[S4].

—

The Cybersecurity Warning

Mistral explicitly states: Mistral Large 4 could help attackers carry out real cyberattacks. Safeguard updates will apply only to the managed API and future releases — not to checkpoints already downloaded. This is a deliberate transparency choice: open weights mean open responsibility. Organizations self-hosting for defensive security (ENISA is testing ML4 for vulnerability identification and code audit ^[S1]) must track their own mitigation pipeline.

—

Where It Fits in the Open-Weight Landscape (November 2026)

Mistral Large 4 is the only Apache/MIT-family licensed trillion-parameter MoE with weights actually on Hugging Face today ^[S1]^[S2]. The MaaS restriction is narrower than Meta’s per-version custom licenses or Qwen’s custom MoE license.

—

Should You Download It?

Download if:

You need a frontier coding agent you can self-host on 4–8 Blackwell GPUs
European data sovereignty matters — trained in EU, EU company, EU infrastructure
You want open weights for cybersecurity tooling (vulnerability discovery, patching, log analysis) with no API dependency
You’re building agentic workflows and need 1M context + native function calling + structured output
You’ll fine-tune on proprietary code/security data via Mistral Forge or LoRA

Wait or skip if:

You don’t have 4× GB300-class GPUs (NVFP4) or 8× GB200 (FP8) — smaller quantizations not yet confirmed for consumer hardware
You need independent benchmark verification before committing infrastructure
Your use case is general chat or creative writing — smaller models (Qwen3.8-27B, DeepSeek V4.1 Flash) are cheaper and sufficient
You’re building a Model-as-a-Service product — the license carve-out requires negotiation
You need vision on non-Blackwell GPUs — NVFP4 targets Blackwell; FP8 vision support on other hardware is unconfirmed

—

What’s Next

Mistral’s roadmap ^[S1]^[S3]:

Upstream vLLM improvements: cross-layer KV sharing, DSpark under DCP/HiSparse, more decode fusions
SkyRL day-0 RL support (LoRA rank 32 on NVFP4, AIME 2024 avg@8 from ~64% → ~75% in 75 steps)
Further distillation for reasoning-effort modes
Forge customization for enterprise security data

The technical report frames the architectural lesson: evaluate on the work the model is meant to support (agentic coding, long-horizon tasks), not language-modeling loss proxies ^[S3]. The KV-cache savings that looked good on perplexity hurt SWE-bench — the fix was restoring context access then making it selective with sparse attention.

—

Bottom Line

Mistral Large 4 is a genuine milestone: the first European-trained, trillion-parameter, open-weight multimodal model with frontier coding/cybersecurity scores and a license that enables sovereign self-hosting for almost everyone except MaaS resellers. The hardware floor is real (4× GB300 for NVFP4), but for teams that clear it, this is the most capable open-weight model you can fully own today.

Download: huggingface.co/mistralai/Mistral-Large-4-1T-A52B (FP8) or mistralai/Mistral-Large-4-1T-A52B-NVFP4 (Blackwell) Serve: vllm/vllm-openai:mistral-large-4 Docker image with day-0 config License: Read the Mistral Ecosystem Open-Weight License before commercial deployment

Sources

[S1] Mistral Large 4 (Le Chonk) is now open weight — Mistral AI (official blog) (2026-10-29)
[S2] mistralai/Mistral-Large-4-1T-A52B · Hugging Face — Hugging Face (model card) (2026-10-29)
[S3] Mistral Large 4 Technical Report — Mistral AI via alphaXiv (2026-11-01)
[S4] Day 0 support for Mistral Large 4 on vLLM — vLLM Blog (2026-10-29)
[S5] Mistral Ecosystem Open-Weight License — Hugging Face (license file) (2026-10-29)

Mistral Large 4 Goes Open Weight: Europe’s First Trillion-Parameter Model

Read aloud using your device