Model

Mistral Large 4

OPEN

Mistral AI · Mistral Large family · released Sep 9, 2026

Mistral's September 9, 2026 flagship — a modest 685B/43B scale-up over Large 3 (675B/41B) that stays Apache 2.0 and doubles context to 512K. The honest read is unchanged from the Large line's recent history: competent open generalist, visibly behind the reasoning-specialized crowd — HLE 9.4 (independent), Terminal-Bench 2.1 at 21.6, SWE-bench 52.7. As with Large 3, the filings here are the independent/third-party measurements Mistral's own docs now link out to rather than vendor table claims; capabilities.reasoning stays false (no thinking mode).

ReasoningCodingVisionFunction callingTool useAgentic

1412.6

Elo · rank #161

Parameters

685B

Active params

43B (MoE)

Context

512K tokens

Architecture

685B-total/43B-active sparse MoE, Apache 2.0 open weights; 512K-token context

License

Apache 2.0

Languages

—

API price (in/out)

$0.6 / $1.8

Modalities

text · vision

Benchmark results

Bar shows position within the tracked field; marker = field best

AIMEMath52.6%#119

best: GPT-5.2 · 100.0%

GPQA DiamondReasoning72.8%#133

best: GPT-6 Astra · 96.0%

Humanity's Last ExamReasoning9.4%#110

best: Claude Opus 5 · 64.7%

HumanEvalCoding91.3%#16

best: Claude Opus 4.5 · 99.4%

IFBenchReasoning44.2%#52

best: Inkling-Medium · 85.3%

LiveCodeBenchCoding62.9%#100

best: DeepSeek-V4.1 (Think Max) · 94.2%

MATH-500Math94.2%#38

best: GPT-5 · 99.4%

MMLU-ProKnowledge78.4%#68

best: Claude Fable 5 · 91.5%

MMMU-ProVision62.7%#51

best: Claude Opus 4.7 · 85.5%

MMMUVision71.8%#50

best: Claude Fable 5 · 89.3%

SWE-bench VerifiedCoding52.7%#100

best: Claude Opus 5 · 96.0%

Terminal-Bench 2.0Coding21.6%#85

best: Gemini 3.8 Flash · 89.4%

Video-MME ProVision71.4%#4

best: Seed 2.2 Pro · 90.4%

Run it locally

VRAM @ Q4

—

VRAM @ FP16

—

Fits on (Q4)

Multi-node cluster required

Throughput data unavailable.

Quantizations

—

Fine-tune it

Permissive

QLoRA465.8 GB4× H200 141GB

LoRA1459.0 GB8× B200 192GB

Full fine-tune10994.3 GBbeyond 8× B200

QLoRA SFT on ~10k samples ≈ $22.24 (4× H200 141GB)

Plan a fine-tune →

Mistral Large family

Elo progression across releases

Mistral LargeFeb 2024620.1 Mistral Large 2 (24.07)Jul 2024984.9 Mistral Large 24.11Nov 2024611.9 Mistral Large 3Dec 20251238.4 Mistral Large 4Sep 20261412.6

Resources

HFmistralai/Mistral-Large-4 DOCShttps://docs.mistral.ai/getting-started/models/models_overview/APIAvailable via Mistral AI API

API price $0.6/$1.8 · each benchmark row carries its own source badge (see methodology)