Mistral Large 4

Mistral has launched a public preview of Mistral Large 4, a one-trillion-parameter multimodal model (code-named "Le Chonk" or ML4) with 49 billion active parameters, trained from scratch on roughly 4,000 Nvidia Grace Blackwell GPUs over about two months in European data centers. The model spans more than 160 languages and targets coding, cybersecurity, finance, manufacturing, and visual grounding workloads. The release follows a staged rollout: a public preview is available now via Mistral's API, followed by a roughly three-week testing period with developers, cybersecurity leaders, and government authorities before open weights publish under a custom Mistral license on October 27. Mistral is positioning ML4 as a foundation enterprises and governments can self-host on sovereign infrastructure with zero-data-retention options, with reinforcement learning still continuing ahead of the final checkpoint.

LLM Stats catalogs Mistral Large 4 benchmark scores sourced from the model's scorecard and official Mistral announcements. Notable reported results include AA-Briefcase 1393/3000, CyBench 93% unguided success rate, SciCode 92% pass@1, and CyberGym 82% on vulnerability reproduce-plus-patch. On cost efficiency the model posts an LLM Stats Score of 46.4 against a blended price of $0.75 per million tokens, placing it between Claude Haiku 5.5 and Claude Opus 5.5 on the aggregator's efficiency curve. Rankings and tier standings across the catalog are listed, though figures are flagged as self-reported by the provider pending independent verification.

BenchLM tracks Mistral Large 4 with a capability score of 53.7/100, ranking 69th of 214 tracked models as of October 6, 2026. The model is listed at $0.68 per million input tokens, $2.09 per million output tokens, and a cached input rate of $0.070, with a blended price of $1.39 per million tokens. Speed measurements show 116 tokens per second with a first-token latency of 18.69 seconds, against a 1 million-token context window (noted as exceeding the field median of 256,000). Benchmark coverage is sparse in this snapshot, with only 3 published rows across the agentic, coding, and reasoning categories and most other categories marked as not measured.

Mistral Large 4 released on October 6, 2026 as an open-weight model from Mistral, ranked 32nd of 44 on the Vals Index with a 48.05% accuracy score. It supports both text and image inputs and offers a 512k context window with up to 256k max output tokens, with input/output token costs listed at $1.36 and $4.18 respectively. Benchmark detail on Vals includes Code Migration 30.56%, EMB 56.21%, Finance Agent v2 54.68%, Legal Research 31.73%, MedCode 40.69%, MedScribe 80.43%, MysteryMechanism 12.16%, ProofBench v1.1 10.00%, and Tax Agent 63.29%, with its strongest result being 6th of 75 on Harvey's Legal Agent Benchmark. Cost per Vals Index test is reported at $13.78 and average latency at 105 minutes 50 seconds.