Mistral Large 4

Mistral Large 4

Quick answer: Mistral Large 4 ("Le Chonk") is Mistral AI's flagship 1-trillion-parameter (52B active) open-weight Mixture-of-Experts model, trained from scratch on European infrastructure. Following its public preview, Mistral shipped the open weights with three weeks of additional RL, taking the lead among open-weight models on FrontierCode and Terminal-Bench 4.0 while matching GLM-5.3 on the Artificial Analysis Coding Agent Index at 47% fewer tokens per task. 1M context, three reasoning modes (low/medium/high).

At a Glance

Where Mistral Large 4 leads

• 48% Artificial Analysis Coding Agent Index — tied with GLM-5.3 (leading open-weight model) at 47% fewer tokens per task
• 41% Terminal-Bench 4.0 — leads open-weight models, +13pp over its own preview release
• 84.5% BrowseComp — ahead of GLM-5.3 and DeepSeek V4.1 Flash on web-browsing research
• Fully open-weight (Mistral Ecosystem Open-Weight License), 1M-token context, native vision input

Where it lags

• Undisclosed pricing — API-served pricing via Mistral not yet published at time of writing
• Single-lab (Mistral) benchmark reporting for several headline numbers (Cyber Index, Finch, B3) — see exclusions below

Best for: Sovereign/self-hosted agentic coding and cybersecurity workloads needing open weights with frontier-tier performance.

What Mistral Large 4 Is

Mistral Large 4 is Mistral AI's largest and most capable model to date: a natively multimodal, ~1T-total/~52B-active-parameter Mixture-of-Experts model (128 routed experts, 4 active per token, plus 1 shared expert), trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters. It unifies instruction following, reasoning, coding, and agentic workflows in one set of weights, with particular strength in cybersecurity, finance, law, and visual grounding.

The model entered public preview in early October 2026; this release ships the open weights after roughly three additional weeks of reinforcement learning, which Mistral reports drove outsized gains in coding and defensive cybersecurity with no signs of saturation. The checkpoint ships an embedded DSpark speculative-decoding draft head (3 dense MLA layers), reported to give up to 1.9x higher decoding throughput.

Three reasoning effort levels (low/medium/high, high by default) let the same weights serve both fast chat replies and long, high-precision agentic runs.

Specifications

Pricing

Open weights: free to download and self-host (hosting/inference cost only). Also served via the Mistral API as mistral-large-4; Mistral has not published per-token API pricing for this release at time of writing — see Mistral's pricing page for current rates.

Context Window

Mistral Large 4 has a 1,048,576-token (1M) context window — roughly 1,500+ pages of text in a single request. Combined with its agentic tool-calling and strict JSON schema enforcement, this supports long-document analysis, large-codebase agentic coding, and multi-document enterprise workflows.

Public Benchmark Scores

Scores above are reported by Mistral AI and shown for context. They are not Benchgen measurements. Figures depend on harness, tools, and effort settings — see the source for methodology.

Excluded from the table above (bucket-C, documented per scout procedure):

• Artificial Analysis Intelligence Index (reported "26") — excluded: the reported value is inconsistent with sibling open-weight entries already on this benchmark's leaderboard (e.g. Meta Muse Spark Open scores 39, a smaller/less-coding-focused model), and Mistral's own framing ("highest among western open-weight models") does not reconcile with that gap. Flagged for follow-up rather than guessed; do not publish without independent confirmation.
• Artificial Analysis Cyber Index (51%) — no local benchmark config exists yet for this index; deferred rather than created hastily under the bucket-B cap. Candidate for a future benchmark-config addition once methodology detail is captured properly.
• FinWorkBench / "Finch" (73.3%) — vendor-named benchmark with no public paper, dataset, or methodology page found; insufficient external reproducibility to add as a benchmark entity.
• B3 Agent Security Benchmark (97.1% attack resistance) — same issue: no public spec/paper located; vendor-internal-style security eval.
• Cognition FrontierCode 2.0 "Main" vs "Extended" subset — the existing frontiercode entity's SOTA table (Grok 4.6, 61.3%) cites the Extended subset; Mistral's 19.7% figure's subset wasn't explicitly stated as Extended in the announcement excerpt reviewed, so it's recorded as the generic FrontierCode score with the announcement as source — flag for a methodology double-check if subset mismatch is later confirmed.

Mistral Large 4 vs Alternatives

Mistral Large 4 is Mistral's first flagship release shipped fully open-weight rather than commercial-only (unlike Mistral Large 3), while matching the leading open-weight coding models at substantially lower token cost per task.

How Mistral Large 4 Performs on Real Agent Tasks

Mistral's reported numbers emphasize token efficiency alongside raw capability: tying GLM-5.3 on the AA Coding Agent Index while using 47% fewer tokens per task materially changes the economics of running coding agents at scale, not just whether a task passes. The BrowseComp and Terminal-Bench gains point to stronger multi-step tool-use and long-horizon terminal task completion versus the October preview checkpoint. As with any single-lab-reported figure, builders should validate against their own harness and task distribution before relying on these numbers for procurement decisions.

Related Models & Benchmarks

• Compare: Mistral Large 3 · GLM-5.3
• Benchmarks: AA Coding Agent Index · FrontierCode · Terminal-Bench 4.0 · BrowseComp

Frequently Asked Questions

Specs and scores sourced from Mistral AI's Mistral Large 4 release announcement (October 2026) and Hugging Face model card. Last updated 2026-10-30.