Mistral Large 4: Europe-Trained MoE Preview at Half List

Mistral Large 4 is a Europe-trained MoE in Studio + API public preview that Mistral positions as open-weight-bound by end of October. It posts strong self-reported coding, agentic, and cyber numbers at a current half-of-list sale on the pricing page. Prefer docs specs: 1.05T total / 52B active / 1.6B vision / 1M context over announce rounding (~1T / 49B).

The honest frame is open-model strength under preview pricing, not a closed-frontier crown. Headline self-reported numbers: 61.7% DeepSWE 1.1, 49.8% Coding Agent Index, 82% CyberGym-E2E, 42.0% Dense200 bbox (vs Astra 41%). Current sale is $0.68 / $2.09 per 1M (list $1.36 / $4.18); no official end date is published. API id: mistral-large-4. Cyber wins partly vs models that refuse. None of these scores are LLM Stats verified yet.

At a Glance

• Catalog / API id: mistral-large-4 (alias mistral-large-4-0 exists; primary is without -0)
• Organization: Mistral AI
• Availability: Studio + API public preview (mistral-large-4). October 6, 2026. Weights planned end of October 2026.
• Architecture: Granular MoE, natively multimodal. Docs: 1.05T total / 52B active + 1.6B vision encoder. Announce rounds to ~1T / 49B.
• Pricing (current sale): $0.68 input / $0.07 cached / $2.09 output per 1M. List: $1.36 / $0.14 / $4.18. No official sale end date published.
• Context: 1,000,000 tokens. Max output not published.
• Modalities: Text + image in, text out
• Training:From scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's European datacenters; preview served on the same infra. Training data multilingual across more than 160 languages, including every official EU language.
• License: Proprietary until the weight drop. Weight license TBD.

What's New

Large 4 is Mistral's frontier open-weight-bound MoE in Studio + API public preview, trained and served in Europe under European law, with weights planned for end of October. That availability story (Studio + API now, weights soon, EU deployment) matters as much as any single bench row.

On capability, Mistral's announce shows a strong open-model showing on coding and agentic harnesses, Dense200 grounding slightly ahead of GPT-6-Astra (42% vs 41%), and cyber numbers where closed models often refuse. SciCode-Verified is claimed as open-weight SOTA at 91.8% pass@1 (n=6).

On price, the current preview sale on the pricing page is half of list: $0.68 / $2.09 versus list $1.36 / $4.18, with 1M context. No official sale end date is published. RL is still in flight per Mistral, so treat scores as a moving preview snapshot.

Benchmarks

All scores below are self-reportedfrom Mistral's announce prose and charts. Not LLM Stats verified. Interpret by job. Do not invent peer scores where Mistral only says "ahead of." Do not claim overall closed-frontier superiority.

Coding and agents

DeepSWE 1.1 at 61.7% and SWE-Atlas-QnA at 59.4% are the software-engineering headlines. Coding Agent Index at 49.8% is ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max per Mistral (no peer percentages published). AutomationBench at 59.9% covers 657 business workflows and is ahead of Kimi K3, MiMo-V2.6-Pro, and DeepSeek V4 Pro. Terminal-Bench 4.0 at 28.3% is the quieter terminal row. AA-Briefcase at 1393 Elo is the long-horizon knowledge-work score, ahead of DeepSeek V4 Pro per Mistral.

Vision, science, and professional

Dense200 bbox at 42.0%edges GPT-6-Astra's 41% on grounding. ChartQA Pro is 63.1%; GDP.pdf is 18.6%. SciCode-Verified pass@1 (n=6) is 91.8% (open-weight SOTA per Mistral). Finance Agent v2 54.7% and Harvey LAB 15.8% are the professional rows; Mistral says finance/legal exceeds Astra and Harvey outperforms all open-source (qualitative peer claims). Finch / FinWorkBench is 67.4%.

Cyber capability

Cybench at 93% (40 CTF-style challenges) and CyberGym-E2E at 82% are the cyber headlines. CyberGym-E2E is highest of any model per Mistral; closed models land near zero partly because they refuse. Read cyber as capability under that harness, not as universal superiority without refusal notes.

Catalog domains by job

Fifteen catalog capability benches, grouped by job. Absolute bars. Safety and refusal rows are skipped. Self-reported; not LLM Stats verified.

How to read the table

Coding agents: DeepSWE, SWE-Atlas, Coding Agent Index, AutomationBench, Terminal-Bench, plus AA-Briefcase Elo for long-horizon work. Vision / docs: Dense200, ChartQA Pro, GDP.pdf. Science / professional: SciCode, Finance Agent, Harvey LAB, Finch. Cyber: Cybench and CyberGym-E2E, with the refusal caveat attached. Prefer Large 4 when Europe-served open-weight-bound pricing and those jobs dominate; prefer closed flagships when you need peak verified score on the hardest suites.

Pricing

First-party Mistral docs, USD per million tokens. Current preview sale on the pricing page is half of list. No official sale end date is published.

Public callers use mistral-large-4. An alias mistral-large-4-0 exists on the docs model page; primary id is without -0.

When to Use / Migrate

• Good fit: Coding and agent loops, business workflow automation, document grounding, finance/legal agent prototypes, and cyber research harnesses where you want a strong open-weight-bound Europe-served option under sale pricing.
• Try now if the current half-of-list sale and 1M context matter and you can accept self-reported Studio + API preview scores while RL is still running.
• Wait for weights if on-prem / self-host license terms matter more than Studio + API preview access. Weights are planned for end of October; license TBD.
• Prefer closed flagships when you need peak verified score on the hardest agentic or science jobs, or when refusal behavior on cyber harnesses is a product requirement rather than a scoreboard artifact.
• Do not migrate on cyber alone. CyberGym-E2E wins partly vs models that refuse. Validate on your own harness before treating 82% as production superiority.

Caveats

• Public preview, not GA. Mistral says RL is still in flight; scores can move.
• All benches above are self-reported by Mistral from announce prose and charts. Not LLM Stats verified. Harness versions and agent scaffolds matter, especially SWE, Terminal, Automation, and cyber.
• Spec conflict: announce rounds to ~1T / 49B active; docs say 1.05T / 52B active + 1.6B vision. Prefer docs for specs.
• Current sale on the pricing page is $0.68 / $0.07 / $2.09 vs list $1.36 / $0.14 / $4.18. No official sale end date is published.
• Cyber wins partly vs closed models that refuse. Do not overread CyberGym-E2E 82% as universal cyber superiority.
• Do not claim Large 4 beats closed frontier overall. Peer leads without published peer percentages stay qualitative.
• Max output and knowledge cutoff are not published on the official model card. Weights / license TBD until the end-of-October drop.
• Safety, refusal, and jailbreak scoreboards are intentionally omitted from this post.

Outlook

Near-term watch items: whether sale pricing changes on the docs page, the end-of-October weight drop and license, and whether independent harnesses confirm the self-reported coding and cyber rows. This post ages when LLM Stats verifies DeepSWE, Coding Agent Index, or CyberGym-E2E, and when weights ship.

Through-line for now: Europe-trained open-weight-bound MoE in Studio + API public preview, strong self-reported open-model showing on coding/agents and cyber at current half-of-list sale, with 1M context and docs-size 1.05T / 52B. For primary sources, see Mistral's Large 4 announcement, the docs model card, and pricing.