Mistral Large 4
Native multimodal MoE with 1.05T total and 49B active parameters, 1M context and 160+ languages. Public API preview is available; weights are planned for the end of October 2026.
At a glance
- Context window
- 1M tokens
- Max output
- —
- Text input / output · per 1M tokens
- $0.700 / $2.10
- Knowledge cutoff
- —
Benchmark snapshot
Best result in each categories
AcademicFrontierMath v2
58.6Thinking High (With Tools) · #46 of 85
Source: Epoch AI (epoch.ai)CodingSWE-Atlas-QnA
59.4Thinking (With Tools) · #4 of 17
General EvaluationLiveBench
71.8Thinking High (No Tools) · #53 of 106
Source: LiveBenchMultimodalMMMU-Pro
76.4Thinking (No Tools) · #55 of 136
Source: Artificial Analysis (artificialanalysis.ai)ProfessionalHarvey's Legal Agent Benchmark
15.8Thinking High (With Tools) · #6 of 52
Science & HealthMedScribe
80.4Thinking High (No Tools) · #53 of 99
Benchmark details
29 results · Best published score per benchmark, grouped by category
| Benchmark · setting | Score | Rank |
|---|---|---|
| Academic | ||
| FrontierMath Tier 4 v2 Thinking High (With Tools) · Mathematics Source: Epoch AI (epoch.ai) | 26.8 | 40 / 73 |
| FrontierMath v2 Thinking High (With Tools) · Mathematics Source: Epoch AI (epoch.ai) | 58.6 | 46 / 85 |
| MysteryMechanism Thinking High (With Tools) · Scientific Reasoning | 12.2 | 32 / 32 |
| ProofBench v1.1 (Lean 4) Thinking High (With Tools) · Mathematics | 10.0 | 49 / 54 |
| Coding | ||
| AA Cyber Index Thinking (With Tools) · Repository Engineering | 49.5 | 7 / 22 |
| Code Migration Thinking High (With Tools) · Maintenance & Optimization | 30.6 | 43 / 79 |
| DeepSWE Thinking (With Tools) · Repository Engineering | 61.7 | 45 / 73 |
| FrontierCode 1.1 Thinking High (With Tools) · Main (100 tasks) · limited comparability · Code Generation & Editing | 39.4 | 30 / 51 |
| FrontierCode 2.0 Thinking High (With Tools + Internet) · Repository Engineering | 8.8 | 30 / 36 |
| IOI (Vals v2) Thinking High (With Tools) · Algorithmic Coding | 45.3 | 32 / 45 |
| SWE-Atlas-QnA Thinking (With Tools) · Repository Engineering | 59.4 | 4 / 17 |
| Terminal-Bench 4.0 Thinking (With Tools) · Agentic Development | 28.3 | 40 / 90 |
| Terminal-Bench 4.1 Max (With Tools) · Agentic Development | 27.6 | 26 / 48 |
| Vibe Code Bench 1-100 Thinking High (With Tools) · Agentic Development | 14.3 | 20 / 25 |
| Vibe Code Bench v1.1 Thinking High (With Tools) · Code Generation & Editing | 78.4 | 36 / 121 |
| WeirdML v3 Thinking High (With Tools) · ML Engineering Håvard Tveit Ihle, WeirdML v3 | 0.0440–1 scale | 25 / 26 |
| General Evaluation | ||
| AA Intelligence Index v5.0.1 Thinking (With Tools) · Capability Indices | 25.3 | 39 / 62 |
| LiveBench Thinking High (No Tools) · Cross-capability Suites Source: LiveBench | 71.8 | 53 / 106 |
| LiveBench 2.0 Thinking High (With Tools) · Cross-capability Suites Source: LiveBench (livebench.ai) | 47.7 | 47 / 51 |
| Vals Index v2.1 Thinking High (With Tools) · Capability Indices | 48.0 | 30 / 38 |
| Multimodal | ||
| MMMU-Pro Thinking (No Tools) · Visual Understanding Source: Artificial Analysis (artificialanalysis.ai) | 76.4 | 55 / 136 |
| Professional | ||
| EMB (Excel Modeling) Thinking High (With Tools) · Finance | 56.2 | 49 / 78 |
| Finance Agent v2 Thinking High (With Tools) · Finance | 54.7 | 24 / 75 |
| Harvey's Legal Agent Benchmark Thinking High (With Tools) · Legal | 15.8 | 6 / 52 |
| Legal Research Bench Thinking High (With Tools) · Legal | 31.7 | 52 / 91 |
| Tax Agent Bench Thinking High (With Tools) · Finance | 63.3 | 14 / 74 |
| Science & Health | ||
| BioMysteryBench Thinking High (With Tools) · Biology & Genomics | 67.0 | 21 / 28 |
| MedCode Thinking High (No Tools) · Clinical Workflows | 40.7 | 58 / 97 |
| MedScribe Thinking High (No Tools) · Clinical Workflows | 80.4 | 53 / 99 |
Rank is this model's position among every model with a published result on that leaderboard; the bar shows how far up the field that is. Orange marks first place on leaderboards with at least 5 models. Unranked leaderboards publish a composite or direction-less score.
Back to benchmark snapshot ↑API pricing
Standard text rates, USD per million tokens · effective 2026-11-13
- Input
- $0.70
- Output
- $2.10
- Cached input
- $0.07
Blended price4:1 input:output · 90% of input tokens cached · same basis as the price ranking
$0.5264/ 1M tokens
Specifications
- Provider
- MistralAI
- Model type
- Multimodal model
- Released
- October 6, 2026
- Total parameters
- 1.1T
- Active parameters
- 49B
- Architecture
- Mixture of experts
- Context window
- 1M tokens
- Input
- Text, Image
- Output
- Text
- Weights
- Closed / API only
- Commercial use
- Proprietary
Capabilities
Editor’s take
LLMLearner Editorial · Figures as of · AI-assisted
Mistral Large 4: a solid European option, not a frontier one
Mistral Large 4, a 1.05T-total / 49B-active multimodal MoE in public API preview, scores 28.3% on Terminal-Bench 4.0 and 63.29% on Tax Agent Bench, and ranks fifth on Harvey’s Legal Agent Benchmark. Its API price is $1.36 / $4.18 per million tokens now that the launch discount has ended; weights are planned for the end of October 2026.
I want to like Mistral Large 4 more than the numbers let me. It is a big, native multimodal MoE (1.05T total parameters, 49B active), it supports 160+ languages and a 1M-token context, and Mistral says the weights are coming at the end of October. For European teams that care where their model comes from, that matters.
But on the benchmarks, it sits in the middle of the pack:
- Terminal-Bench 4.0: 28.3% (26th of 65), behind the open-weight DeepSeek-V4.1-Flash-1017 (42.3%) and MiniMax M3.1 Flash (38.1%).
- Vibe Code Bench v1.1: 78.4% (28th of 111).
- MMMU-Pro: 76.42% (48th of 126).
- FrontierMath v2: 58.6% (42nd of 81). Lean 4 proofs are a clear weakness at 10% on ProofBench v1.1.
There are bright spots in professional work: fifth on Harvey’s Legal Agent Benchmark (15.83%), fourth on SWE-Atlas-QnA (59.4%), fifth on the AA Cyber Index (49.51) and 63.29% on Tax Agent Bench, a score well above several more expensive models.
On price: the 50%-off launch window (two weeks from October 6) has ended, and the list price is now $1.36 input / $4.18 output per million tokens, with cached input at $0.14. The DeepSeek and MiniMax flash models score higher on most agentic boards for a fraction of that.
My take: worth evaluating for legal, tax and multilingual European workloads, and worth revisiting when the open weights land. As a general coding or agent model, the open Chinese flash models are stronger and cheaper.
Model analysis
Mistral Large 4, also called Le Chonk, is Mistral AI’s native multimodal MoE model, released as a public API preview on October 6, 2026. Official documentation lists 1.05T total parameters, 49B active parameters and a 1.6B vision encoder. It accepts text and images and generates text, combining instruction, reasoning and agent capabilities. Reinforcement learning is still running; weights are planned for the end of October. The preview is therefore an intermediate release.
Official specifications now include 1M context and 160+ languages
The model documentation specifies a 1M-token context window, function calling, structured outputs and tool use; its example uses high reasoning effort. The release page describes native fluency in 160+ languages and training data covering every official EU language, without publishing a complete supported-language list. Mistral says training took place from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European datacenters, which also serve the public preview. Supervised fine-tuning and ongoing reinforcement learning cover scientific reasoning, coding, safety and long-horizon tool use.
Vals lists 512K context, a 256,000-token output limit, temperature 1, top-p 0.95 and high reasoning effort. Those are evaluator settings, not replacements for the official specifications. An official maximum output length and knowledge cutoff remain undisclosed.
Independent professional results coexist with private AA coding evaluations
As captured on October 6, Artificial Analysis reports 59.9% on AutomationBench-AA and 49.51 on its Cyber Index, rounded to 50 in the launch chart. Vals reports 54.68% on Finance Agent v2, 15.83% on Harvey’s Legal Agent Benchmark and 48.05% on Vals Index v2.1. These measure distinct task sets and provide evidence for particular professional and security capabilities. They do not, individually, establish Mistral’s broader claim of leadership among US or European open-weight models.
The official release also reports 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA and 28.3% on Terminal-Bench 4.0. Its footnote attributes these to private Artificial Analysis evaluations ahead of a new harness’s public launch, with inclusion in the Coding Agent Index planned when that harness is released. The harness and reasoning budget have not been published. The 28.3% result must remain separate from Vals’s 22.73% result using Mini-SWE-agent. Mistral also reports a 49.8% Coding Agent Index without identifying its version.
The visual-grounding lead is evidence from a specific test
Mistral reports 42% on Dense 200 against GPT-6 Astra’s 41%, a one-percentage-point difference in its own evaluation. That does not establish superiority across all visual tasks. Its demonstrations cover engineering drawings, PDF evidence retrieval and large satellite images. The company reports 93% on Cybench and 93.3% attack resistance on Lakera B3, without full evaluation settings. Its vetted cybersecurity partners also test the same model with reduced moderation; that setting should be distinguished from the public API.
Launch pricing is discounted; self-deployment awaits weights and a license
The live documentation offers a two-week 50% launch discount: $0.68 input, $0.07 cached input and $2.09 output per million tokens, compared with original prices of $1.36, $0.14 and $4.18. The release-page footer and Vals use original prices. The public preview can be used through Mistral’s API today. Teams seeking self-deployment must wait for weights and licensing details, both still marked Coming soon in the official documentation.
Sources: Mistral release article; official model documentation; Artificial Analysis model page; AA Cyber Index; Vals model evaluation. Accessed October 6, 2026.
Last reviewed 2026-10-06
Mistral Large 4 vs competitors
Comparable results only · record is Mistral Large 4's wins–losses
| Benchmark | Mistral Large 4This model | vs GPT-6 AstraRecord 4–23 | vs Claude Opus 5.5Record 2–25 | vs Gemini 3.8 FlashRecord 7–20 | vs Claude Fable 5.1Record 2–24 |
|---|---|---|---|---|---|
| MMMU-Pro | 76.4 | 87.8 | 89.8 | 85.6 | 88.4 |
| Vibe Code Bench v1.1 | 78.4 | 89.6 | 90.3 | 78.7 | 90.3 |
| LiveBench | 71.8 | 82.2 | 83.2 | 75.8 | 83.4 |
| MedScribe | 80.4 | 87.9 | 91.4 | 84.5 | 91.3 |
| MedCode | 40.7 | 48.5 | 49.8 | 48.1 | 53.5 |
| Legal Research Bench | 31.7 | 39.4 | 50.5 | 38.9 | 55.3 |
| FrontierMath v2 | 58.6 | 93.7 | 91.2 | 68.4 | 90.2 |
| Code Migration | 30.6 | 67.7 | 66.7 | 36.5 | 54.6 |
| EMB (Excel Modeling) | 56.2 | 71.7 | 75.9 | 72.2 | 76.7 |
| Finance Agent v2 | 54.7 | 53.5 | 58.6 | 61.4 | 58.9 |
Primary sources
- Mistral Large 4 technical reportmistral.ai
- Mistral Large 4 online democonsole.mistral.ai
Community discussion
Share the API, setup, and tradeoffs you actually tested. Rate the model once; comment as often as you like.
Used this model? Sign in with Google to rate it and share what worked, what did not, and which API or setup you tested.
No community comments yet.
