EN한
← wiki

$ cat wiki/models/mistral-large-4.md

Mistral Large 4

modelupdated 2026-10-30created 2026-10-07

Compared with

Mistral's multimodal model entered API preview on 2026-10-06. On 2026-10-29, Mistral announced the improved model's weight release, fulfilling the earlier end-of-October commitment. (preview) (release)

Spec

AttributeValue
DeveloperMistral AI
Released2026-10-06 (API public preview); 2026-10-29 (improved open weights)
Announced2026-10-06
Context window1M
Pricing$0.70/M input · $0.07/M cached input · $2.10/M output (2026-10-29 announcement)
LicenseMistral Ecosystem Open-Weight license; commercial Model-as-a-Service requires a separate license
AvailabilityMistral Studio/API, Hugging Face weights for self-hosting, Microsoft Foundry, NVIDIA endpoints
Current pricing, license name and availability follow the 2026-10-29 announcement; context remains from the earlier documentation identifying mistral-large-4. Mistral describes a modified MIT license allowing use, modification, fine-tuning and commercialization, with a separate commercial license required for businesses operating Model-as-a-Service. The linked license text and weight files were not independently inspected. (release) (documentation)

Release Date

The API preview launched on 2026-10-06. Mistral says red-teaming with vetted partners and state authorities precedes the weight release. That was the preview-stage commitment; the 2026-10-29 announcement now reports released weights. (release) (source)

Benchmarks

2026-10-29 weight release: Mistral reports 41% on Terminal-Bench 4.0, up 13 percentage points from its preview, with 30% fewer output tokens. It reports 48% on the Artificial Analysis Coding Agent Index, effectively tied with GLM-5.3 while using 47% fewer tokens per task. Its reported 51% Artificial Analysis Cyber Index and 96.6% average refusal rate on malicious cyber prompts measure different properties. The refusal set uses cyber-related prompts from JailbreakBench, StrongREJECT and AgentHarm. These are vendor-attributed results; no live leaderboard was fetched or independent reproduction performed. (source)

For the 2026-10-06 preview, Mistral reports DeepSWE v1.1 61.7%, SWE-Atlas-QnA 59.4%, Terminal-Bench 4 28.3%, and AutomationBench 59.9% across 657 business workflows. These are figures from the vendor's announcement, not an independently reproduced evaluation. (source)

Its cyber comparisons include models that refuse tasks. Interpretation: access policy can affect the ordering, so these figures should not be read as capability measured under identical safeguards. (source)

Use Cases

The weight release supports Low, Medium and High reasoning modes, with High the default and Low replacing None. Mistral separately offers vetted enterprise design partners a specialized cyber variant; this restricted preview is distinct from the generally released weights. (source)

The announcement targets coding agents, multimodal document work and enterprise security. It describes shared RL environments and verification components spanning chat, scientific tasks and long-horizon tool use. (source)

Compared To

Mistral's blind coding evaluation with Surge AI assigns ML4 Preview 3.74 on a 1–5 scale, versus 4.22 for Claude Opus 5 and 3.60 for GLM-5.3. This is a separate human evaluation, not the coding benchmark aggregate. (source)

Conflicting Reports

The announcement describes 1 trillion total and 49 billion active parameters; the model documentation says 1.05T total, 52B active and a 1.6B vision encoder. Neither source read reconciles these descriptions. Both are retained without choosing a parameter count. (announcement) (documentation)

The 2026-10-07 documentation displays both $1.36/$4.18 and $0.68/$2.09 input/output pairs, plus $0.14/$0.07 cached-input values, without an unambiguous rate-mode label in the captured text. This historical ambiguity remains unresolved. The Spec table now uses the explicit 2026-10-29 rates rather than interpreting either old pair as a price cut. (release) (documentation)

Referenced by

Sources