Mistral Large 4

Mistral·2026-10-06

LLMsopen-weightcloud

Overview

1.05T-parameter MoE (~490B active, plus a 1.6B vision encoder), text and image input, text output, nicknamed "Le Chonk". Announced at AI Everything Abu Dhabi on 2026-10-06 with an API preview; open weights scheduled for 2026-10-27. Mistral states a context of up to 1M tokens, OpenRouter serves 512k. Mistral published no GPQA, HLE, MMLU-Pro or SWE-bench Pro at launch; its table leads with DeepSWE v1.1 plus domain evals (FinWorkBench, Harvey Legal Agent Benchmark, visual grounding).

Capabilities and innovations

1.05T total / ~490B active MoEText + image inputUp to 1M token context (per Mistral)Agentic codingOpen weights scheduled 2026-10-27Largest Mistral model to date

Benchmarks

Benchmark	Score	Source
Terminal-Bench 4.0	22.73%	independent Terminal-Bench 4.0 22.73% (±0.88), measured by Vals AI (accessed 2026-10-06). Mistral reports 28.3; the gap exceeds 3 points, so the independent value is primary.
DeepSWE v1.1	61.7%	vendor DeepSWE v1.1 61.7 from Mistral's launch table (headlines round it to 62). Mistral's own model, explicitly v1.1. The same table lists GLM-5.3 at 61, DeepSeek V4 Pro 57, Qwen 3.8 Max 51 and Reflection Beam 44; those are Mistral's measurements of other vendors' models and are not taken over. Read via secondary coverage, mistral.ai was not reachable from the intake sandbox.

GPQA Diamond: not reported by the vendor
Humanity’s Last Exam (no tools): not reported by the vendor
SWE-bench Pro: not reported by the vendor
MMLU-Pro: not reported by the vendor
Terminal-Bench 2.x: not reported by the vendor
MMMU-Pro: not reported by the vendor

API pricing

$0.68 per million input tokens, $2.09 per million output tokens (USD, cheapest OpenRouter endpoint, checked 2026-10-06)

Architecture and hardware

Parameters: 1050B total, 490B active per token (MoE)

More from Mistral

Data curated by AI Model Timeline. See the methodology for admission criteria, benchmark eras and source priority.