Mistral Large 4: Inside Europe's First 1-Trillion-Parameter Open-Weight Model

Mistral AI released a public preview of Mistral Large 4 on October 6, 2026 — a 1.05-trillion-parameter multimodal Mixture-of-Experts model that the company has nicknamed "Le Chonk." The weights are promised by the end of October, making it the largest open-weight model to come out of a European lab. This post walks through the architecture, training methodology, benchmark results, and what the release means for practitioners who want to run frontier-scale models on their own infrastructure.

Architecture: Granular MoE with a Vision Encoder

Mistral Large 4 is a sparse Mixture-of-Experts model with 1.05 trillion total parameters and 52 billion active parameters per token. The "granular" MoE design differs from coarser expert layouts: instead of routing tokens to a small number of large expert blocks, the router selects from a larger pool of smaller, more specialized units. This gives the routing mechanism finer-grained control over which capacity is applied to each token, which Mistral says improves both throughput and task specialization.

A separate 1.6-billion-parameter vision encoder handles image inputs. The encoder converts visual content — charts, engineering drawings, satellite imagery, PDFs — into representations that the language backbone can reason over. The result is a single model that handles instruction following, multi-step reasoning, and multimodal understanding without switching between separate systems.

The model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European data centers, not on rented cloud capacity. The context window is listed as 1 million tokens in Mistral's documentation, though the preview API currently enforces a 524,000-token limit on some third-party providers.

Reinforcement Learning at Scale

Mistral describes its post-training approach as "RL at scale" — a reinforcement learning pipeline that adapts dynamically as the model's capabilities change during training. Rather than running a fixed RL curriculum, the system continuously adjusts the task distribution to stay at the edge of the model's current ability.

At the current compute scale (roughly 3,000 GPUs), a single training run produces approximately 33 billion tokens per day, of which around 16 billion are trainable completion tokens after filtering and masking. The pipeline is optimized for long trajectories, supporting rollout budgets of millions of tokens across multiple compaction steps — which is what allows the model to learn from complex, multi-turn agentic tasks rather than just short-form completions.

Importantly, Mistral notes that the RL run behind the preview is still in flight. The model is expected to continue improving before the weights ship, which is an unusual transparency about the release timeline.

Benchmark Results

Mistral published results across several domains. Here are the headline numbers from the official announcement:

Agentic coding:

• DeepSWE v1.1: 61.7%
• SWE-Atlas-QnA: 59.4%
• Terminal-Bench 4: 28.3%
• Combined Coding Agent Index: 49.8% (ahead of DeepSeek V4 Pro and Qwen3.8 Max)

Agentic workflows:

• AutomationBench (657 business workflows): 59.9%
• AA-Briefcase (long-horizon knowledge work): 1,393 Elo

Cybersecurity:

• Cybench (40 CTF-style challenges): 93%
• Vulnerability reproduction and patching test: 82%
• B3 AI Security Benchmark (attack resistance): 93.3%

Multimodal:

• Dense 200 visual grounding: 42% (vs. 41% for GPT-6 Astra)

In a blind human evaluation run with Surge AI, professional annotators rated the preview 3.74/5, placing it second among five models — behind Claude Opus 5 (4.22) and ahead of GLM-5.3, Kimi K3, and GLM-5.2.

Independent evaluator Artificial Analysis scores the preview at 38 on its Intelligence Index (with reasoning enabled), up from 9 for the non-reasoning Mistral Large 3. Open-weight models from Chinese labs still score higher on that index — GLM-5.3 at 45, Kimi K3 at 44 — but Mistral Large 4 leads among open-weight models developed outside China.

Cybersecurity as a First-Class Use Case

One of the more deliberate design choices in Mistral Large 4 is its treatment of cybersecurity as a primary capability rather than an afterthought. The model is being red-teamed by cybersecurity organizations and government authorities before the weights are released, and Mistral is offering a version with reduced moderation and expanded cyber capabilities to vetted defenders.

The rationale is practical: in incident response and vulnerability research, provider-level refusals can block legitimate work. A model that can be self-hosted under an organization's own policies — and that won't refuse to analyze malware or reproduce a known CVE — is meaningfully different from a hosted API with conservative safety filters. The Artificial Analysis Cyber Index places Mistral Large 4 in the global top five, and it leads open-weight models developed outside China on that benchmark.

Pricing and Self-Hosting

The preview API is priced at $1.36 per million input tokens and $4.18 per million output tokens (with a 50% batch discount). Mistral also offers a European regional endpoint at api.eu.mistral.ai for organizations with data residency requirements, at 1.1× the standard list price. A detailed pricing breakdown is available in Mistral's API documentation.

Self-hosting will require significant hardware. At 8 bits per parameter, 1.05 trillion parameters occupy roughly 1 TB of memory before KV cache and runtime overhead. That puts it in the range of a multi-node GPU cluster for full-precision inference, though quantized deployments will reduce that substantially. Mistral has not yet announced a license for the weights.

What This Means for Practitioners

A few things stand out about this release:

The open-weight commitment matters. Most 1T-parameter models are closed APIs. Mistral Large 4 being open-weight — even if the license is not yet announced — means organizations can audit the model, fine-tune it, and run it on private infrastructure. That is a different risk profile than a hosted API.

The RL training pipeline is the real story. The benchmark numbers are competitive, but the more interesting detail is the training infrastructure: 33 billion tokens per day, long-trajectory rollouts, and a pipeline that adapts to the model's current capability level. That is the kind of post-training machinery that separates frontier labs from everyone else, and Mistral is building it in Europe.

Multimodal is now table stakes. The vision encoder is not a separate product — it is part of the base model. For practitioners building document processing, satellite imagery analysis, or engineering workflow tools, having a single model that handles both text and images simplifies the stack considerably.

The hardware bar for self-hosting is high. 1 TB of memory at 8-bit precision is not a single-server deployment for most organizations. Quantized versions will help, but practitioners should plan for multi-node setups or managed inference until smaller distilled variants appear.

Conclusion

Mistral Large 4 is a technically serious release: a 1-trillion-parameter MoE with a granular routing architecture, a dedicated vision encoder, and a reinforcement learning pipeline designed for long-horizon agentic tasks. The benchmark results are competitive with the current frontier, and the open-weight commitment — unusual at this scale — gives organizations a path to sovereign deployment that closed APIs cannot offer.

The weights are expected by the end of October 2026. Until then, the preview API is available through Mistral Studio, and the full architecture details and post-training methodology will accompany the weight release. For teams evaluating frontier open-weight models for agentic coding, cybersecurity, or multilingual enterprise workflows, this is worth tracking closely. Weights will be published on Mistral's HuggingFace page.