Mistral shipped Large 4 yesterday, internally nicknamed "Le Chonk." A mixture-of-experts with 1.05T total parameters and 49B active per token, natively multimodal, supporting 160+ languages. Available now via Mistral's API. The weights land October 27, after a three-week review with government authorities and cybersecurity partners.

The coding benchmark is third among open-weight models, and Mistral isn't hiding it. DeepSWE v1.1 comes in at 61.7%, behind both Kimi K3 and DeepSeek V4.1 Flash per Mistral's own disclosure. Terminal-Bench 4.0 sits at 28.3% — in roughly the same tier as Grok 4.7 and well below Claude Sonnet 5.5's 70.6% on the same benchmark. The multimodal story is better: Mistral claims Large 4 outperforms some closed frontier models on visual grounding and dense document tasks, though no methodology page has been linked yet.

Source spread

• Mistral — Large 4 announcement — [hype]. First-party launch post; all official specs, benchmarks, and open-weight timeline sourced here.
• VentureBeat — Mistral debuts Large 4 'Le Chonk' — [builder]. Developer-focused coverage, open-weights implications, October 27 confirmation.
• Stork.ai — Le Chonk: benchmarks and trade-offs — [skeptic]. Detailed comparison showing Large 4 third among open-weight models on coding benchmarks vs. Chinese competition.
• MarktechPost — Mistral AI Releases Mistral Large 4 — [builder]. Architecture specs and benchmark breakdown.

Pros & cons

What's real:

• The scale is genuine. 1.05T total parameters with 49B active per token is a proper large MoE — not a numbers game. Per-token compute is efficient; the trained representation is massive. This is the largest open-weight model Mistral has shipped.
• Natively multimodal at this size is a real product story. The visual grounding claim — beating some closed frontier models on dense document and diagram comprehension — is the most interesting angle in the launch and the one worth testing first.
• A firm October 27 date for open weights, with named reviewers (government authorities, cybersecurity partners), is more process discipline than most open-weight drops.
• 160+ language support at frontier scale matters enormously for European and multilingual enterprise deployments. Mistral is genuinely the strongest option in this category.

What deserves a side-eye:

• Third on DeepSWE v1.1 is not leading for coding. Kimi K3 and DeepSeek V4.1 Flash have higher published scores, per Mistral's own disclosure. If coding agents are your primary workload, route traffic there instead.
• Terminal-Bench 4.0 at 28.3% puts Large 4 in roughly the Grok 4.7 tier. Well behind Sonnet 5.5 (70.6%) for autonomous terminal tasks.
• The open-weight license hasn't been announced. "TBD" from a company that has shipped custom commercial licenses before is not automatically reassuring. The October 27 license announcement is more important than the weights themselves.
• At $1.36/M input, you're paying more per token than DeepSeek V4.1 Flash ($0.435/M) for a model that scores lower on the main coding benchmarks. The value case is multimodal, multilingual, and self-hosting once the weights land.

Samwise's take

What builders need to know

Further reading

• Mistral — Large 4 announcement — official specs, benchmark framing, open-weight timeline
• VentureBeat — Mistral debuts Large 4 'Le Chonk' — developer-focused coverage
• MarktechPost — Mistral AI Releases Mistral Large 4 — detailed specs and benchmarks
• Stork.ai — Le Chonk benchmarks and trade-offs — critical analysis of benchmark position vs. Chinese open models
• OpenRouter — DeepSeek V4.1 Flash pricing — cost comparison context