Model
Model explorer

Mistral Large 4

OPEN
Mistral AI · Mistral Large family · released Oct 6, 2026

Mistral's Oct 6, 2026 public preview of its largest model to date: a 1.05T-total/52B-active MoE with a 1.6B vision encoder and 1M context, trained on 3,800 Grace Blackwell GPUs in Mistral's own European datacenters. The preview API is priced at $1.36/$4.18 per Mtok and weights are promised by the end of October, so no license or Hugging Face repo exists yet. Mistral reports DeepSWE v1.1 61.7%, SWE-Atlas-QnA 59.4%, Terminal-Bench 4 28.3% and AA-Briefcase 1,393 Elo, plus strong cyber numbers (Cybench 93%). Most launch figures appear only in prose; the charts are images.

ReasoningCodingVisionFunction callingTool useAgentic
2441.3
Elo · rank #50
Parameters
1050B
Active params
52B (MoE)
Context
1M tokens
Architecture
1.05T-total/52B-active granular MoE with a 1.6B vision encoder; 1M-token context; public preview, open weights promised by end of Oct 2026
License
Not yet published (open weights promised by end of October 2026)
Languages
—
API price (in/out)
$1.36 / $4.18
Modalities
text · vision
Benchmark results
Bar shows position within the tracked field; marker = field best
Arena EloHuman preference1431#60
best: Muse Wonder · 1513
DeepSWECoding61.7%#24
best: Muse Wonder Full · 79.0%
best: Muse Spark 1.2 · 25.4%
best: Step 5 Preview · 63.6%
Terminal-Bench 4.0Agents27.8%#24
best: Claude Sonnet 5.5 · 70.6%
Vals Finance AgentAgents54.7%#22
best: Gemini 4 Argon · 65.4%
Run it locally
VRAM @ Q4
—
VRAM @ FP16
—
Fits on (Q4)
Multi-node cluster required
Throughput data unavailable.
Quantizations
—
Fine-tune it
Conditional / custom
QLoRA714.0 GB4× B200 192GB
LoRA2236.5 GBbeyond 8× B200
Full fine-tune16852.5 GBbeyond 8× B200
QLoRA SFT on ~10k samples ≈ $18.59 (4× B200 192GB)
API price $1.36/$4.18 · each benchmark row carries its own source badge (see methodology)