Atlas

Models

← All models

Mistral Large 4 Preview

Mistral AI · 2026-10-06

Mistral Large 4 is the multimodal mixture-of-experts public preview released on October 6, 2026. The hosted model accepts text and images and supports high or none reasoning effort. Mistral's model documentation lists a one-million-token context window. Weights are announced for later in October and were not yet released at this snapshot. Parameter counts are left unspecified because the launch announcement reports 1T total and 49B active while the model documentation reports 1.05T total and 52B active plus a 1.6B vision encoder.

Benchmark scores

BenchmarkScore
A-Fantasia (Valid Attempts with Retries)65.7
A-Fantasia Chess (Valid Attempts with Retries)61.0
A-Fantasia Cube (Valid Attempts with Retries)70.0
A-Fantasia Spelling (Valid Attempts with Retries)66.0
AA-Briefcase v1.1 Elo1393
AA-LCR v1.181.3
AA-Omniscience Index-5.3
Arena Code WebDev — Brand & Marketing1539
Arena Code WebDev — Consumer Product1554
Arena Code WebDev — Data & Analytics1459
Arena Code WebDev — Gaming1567
Arena Code WebDev — HTML1471
Arena Code WebDev — React1551
Arena Code WebDev — Reference-Based Design1570
Arena Code WebDev — Simulations1541
Artificial Analysis Intelligence Index v4.3.238.4
Artificial Analysis Omniscience Accuracy25.8
Artificial Analysis Omniscience Hallucination Rate41.9
AutomationBench-AA 1.0.659.9
BioMysteryBench — Hard (Vals)29.4
BioMysteryBench — Human-solvable (Vals)75.8
BioMysteryBench (Vals Terminus 2)67.0
Capability158.9
Code Migration30.6
Code Migration - CLI17.4
Code Migration - COBOL70.0
Code Migration — Code Quality4.1
Code Migration (Vals Index v2 subset)22.8
CritPt (avg@5)10.6
Eval Connections One-Shot Group Accuracy83.8
Loading Atlas data…