Mistral Large 4 open weights: Europe's 1T MoE is now downloadable
Mistral AI's Mistral Large 4 — internally nicknamed 'le Chonk' — has shipped its open weights in late October 2026, following an October 6 public preview. The 1.05 trillion parameter natively multimodal MoE (52B active) was trained on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's European datacenters. It leads Western open-weight models on cybersecurity (AA Cyber Index top 5, 82% reproduce-and-patch, 93% Cybench) and competes on coding (61.7% DeepSWE) and agentic workflows (59.9% AutomationBench). Released under Apache 2.0 license.
- Open weights released late October 2026 — Mistral committed to end-of-October release; weights landed on Hugging Face (mistralai/Mistral-Large-4-1T-A52B) with Apache 2.0 license. Preview API was live since October 6 on Mistral Studio and Mistral API (model ID: mistral-large-4).
- 1.05T total, 52B active, natively multimodal — Mixture-of-Experts architecture with vision encoder. 512K context, 256K max output. Trained from scratch on 3,800 Grace Blackwell GPUs in European datacenters under European law — a sovereign AI milestone.
- Cybersecurity lead among Western open weights — AA Cyber Index: top 5 globally, leads open weights outside China. Reproduce-and-patch (CyberGym-E2E): 82%, highest of any model tested. Cybench (40 CTF challenges): 93%. Closed models (Opus 5.5, GPT-6 Astra) score near zero due to safety refusals on defensive cyber tasks.
This article is based on Mistral's official announcement (October 6, 2026), Mistral Docs model page (v26.10), The New Stack coverage (October 2026), Reuters reporting, Artificial Analysis independent benchmarking, and Local AI Zone's October 2026 model updates timeline. The open weights release date is reported as end of October 2026; exact day and final license confirmed as Apache 2.0 per Mistral's announcement. No hands-on testing was performed for this article.
On October 6, 2026, Mistral AI announced Mistral Large 4 — a 1.05 trillion parameter natively multimodal Mixture-of-Experts model, trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in the company's own European datacenters. The public preview went live immediately on Mistral Studio and the Mistral API. Mistral committed to releasing the open weights by end of October 2026.
As of late October 2026, the weights have landed on Hugging Face (mistralai/Mistral-Large-4-1T-A52B) under the Apache 2.0 license. This makes ML4 the largest open-weight model from a Western lab, and the first trillion-parameter European model available for sovereign self-hosting.
What Mistral Large 4 actually is
| Spec | Value |
|---|---|
| Architecture | Natively multimodal MoE (hybrid instruct + reasoning) |
| Total parameters | 1.05 trillion |
| Active parameters | 52B per token |
| Context window | 512K tokens |
| Max output | 256K tokens |
| Input modalities | Text, images (documents, charts, natural images, geospatial) |
| Training compute | 3,800 Grace Blackwell GPUs, European datacenters |
| License | Apache 2.0 |
| Model ID (API) | mistral-large-4 / mistral-large-4-0 |
The model unifies instruction following, reasoning, and agentic capabilities in a single checkpoint — no separate "thinking mode" toggle needed. Mistral's RL library combines single-turn chat, complex problem solving, safety alignment, factuality, and long-horizon tool use in shared training runs producing ~33B tokens/day (16B trainable).
Benchmark claims — vendor + independent
| Benchmark | ML4 Score | Context |
|---|---|---|
| AA Cyber Index | Top 5 globally | Leads open weights outside China |
| Reproduce & patch (CyberGym-E2E) | 82% | Highest of any model tested |
| Cybench (40 CTF challenges) | 93% | One of highest for open weights |
| DeepSWE v1.1 | 61.7% | Ahead of DeepSeek V4 Pro, Qwen3.8 Max |
| Terminal-Bench 4.0 | 28.3% | Competitive coding agent test |
| AutomationBench (657 workflows) | 59.9% | Ahead of Kimi K3, DeepSeek V4 Pro |
| AA-Briefcase (knowledge work) | 1,393 Elo | Ahead of DeepSeek V4 Pro |
| Dense 200 (visual grounding) | 42% | Surpasses GPT-6 Astra (41%) |
Figures from Mistral's October 6 announcement and Artificial Analysis independent evaluation. Blind human evaluation (Surge AI) on coding quality ranked ML4 Preview second of five models (3.74/5), behind only Claude Opus 5 (4.22).
Why the cybersecurity angle matters
The standout claim is cybersecurity performance. Mistral argues that several leading closed models — including Claude Opus 5.5 and GPT-6 Astra — score near zero on reproduce-and-patch tests because their safety filters refuse to perform the task. Yet defending software often starts with proving a flaw is real — exactly the work those refusals block. ML4 is designed to do that work without the same constraints, and its open weights mean organisations can run it on private cloud or on-premise for sovereign security operations.
This is a structural differentiator: open weights + no refusals on defensive cyber tasks is a combination no other Western frontier lab currently offers. For organisations that need auditable, sovereign AI for security operations, this is the primary reason to track the weights release.
European sovereignty and infrastructure
ML4 was trained from scratch on Mistral's own European infrastructure — 3,800 Grace Blackwell GPUs in datacenters the company operates end-to-end under European law. The model will be available across multiple regions including a European deployment independent of other digital service providers. This matters for organisations with data sovereignty requirements, particularly in finance, defence, and public sector.
The €3 billion Series D (largest European tech equity round ever) is funding compute scale-up. Mistral states the RL run behind this preview is still in flight with no signs of saturation, implying further improvements before and after the weights drop.
What to watch
- Independent reproduction: Now that weights are public, expect third-party benchmarking within days. The Cyber Index and DeepSWE claims are the ones to verify first.
- Hardware requirements: 1.05T parameters ≈ 525 GB at INT4, 1.05 TB at FP8, 2.1 TB at BF16. Single-node production needs 8×B200 or equivalent; multi-node or quantized serving (AWQ, GPTQ) will be the practical path for most teams.
- Specialised descendants: Mistral states ML4 will serve as the base for a new generation of specialised models for finance, manufacturing, cybersecurity, and other verticals.
- License verified: Apache 2.0 confirmed — commercial use permitted with attribution.
For model evaluators: The 1T/52B MoE architecture at open weights is a new point on the Pareto frontier. Compare against Qwen3.8-2.4T-A95B (open, larger) and DeepSeek V4 Pro (open, cheaper API) on your actual workloads.
For the open ecosystem: A European lab shipping a trillion-parameter open model trained on European compute is a milestone for AI sovereignty. The weight drop makes this the most significant major open release of Q4 2026.
Bottom line: Mistral Large 4 open weights are now downloadable under Apache 2.0. The cybersecurity lead among Western open weights is the strongest claim — if it holds on independent reproduction, ML4 becomes the default choice for organisations that need open, sovereign, refusal-free defensive AI. The 1T/52B MoE architecture at Apache 2.0 is a new Pareto point for self-hosted AI. Verify benchmarks against your workloads, but the weight drop itself is a milestone for European AI sovereignty.
- Mistral — Introducing Mistral Large 4 (official announcement, October 6, 2026)
- Mistral Docs — Mistral Large 4 model page (v26.10)
- Reuters — Mistral Large 4 coverage (October 2026)
- The New Stack — Mistral Large 4: weights by end of October
- Artificial Analysis — Mistral Large 4 provider benchmarking
- Local AI Zone — October 2026 AI Model Updates: Mistral Large 4 weights
- Hugging Face — mistralai/Mistral-Large-4-1T-A52B (weights release)
Prices and benchmark figures are standard published rates as of October 28, 2026 and can change without notice. Some outbound links may be affiliate links — see our disclosure. Found an error? Tell us and we'll correct it in public.