Back to the index
Mistral AI · Mistral
No. 20

Mistral Large 3

Europe's flagship open-weight MoE — a strong, cheap, multilingual generalist.

OpenOpen weightsText, Image, Pdf
Field notes
Context window
262Ktokens
Max output
33Ktokens
Input price
$0.50/ 1M tokens
Output price
$1.50/ 1M tokens
Modalities
Text, Image, Pdf
Knowledge cutoff
—
Released
Dec 2025
Parameters
675B
License
Apache 2.0
Approx. speed
—
Overview

What this model is

Mistral Large 3 (2512) is Mistral's open-weight flagship, a 675B-parameter sparse MoE with 41B active parameters under a permissive Apache 2.0 license. It pairs a 256K context, text+image input, and best-in-class European-language coverage at a low price. It is genuinely excellent value for multilingual and enterprise work, but sits below the Claude/GPT frontier on long-horizon agentic coding reliability.

Strengths

  • Best-in-class multilingual (esp. European languages)
  • Fully open weights under Apache 2.0
  • Aggressive pricing for a flagship-class model
  • 256K context with vision input

Trade-offs

  • Agentic coding reliability below the frontier
  • Optimized for breadth over deep reasoning
  • Vision capability is secondary
  • Large to self-host despite sparse activation

Best for

  • Multilingual and European-market apps
  • Enterprise/on-prem deployments
  • General reasoning and writing
  • Data-residency-constrained workloads

Not ideal for

  • Long autonomous coding agents
  • Hardest math/science reasoning (use Magistral)
Capabilities

Capability profile

Normalized 0–100 scores, comparable across the whole catalog.

Reasoning74Coding70Math72Writing78Knowledge80Speed66Agentic66Vision60Multilingual86Long Context76
Reasoning
74
Coding
70
Math
72
Writing
78
Knowledge
80
Speed
66
Agentic
66
Vision
60
Multilingual
86
Long Context
76
Benchmarks

How it scores

Public benchmark results, with independent third-party results where available. Bars normalize percentages to 100 and Elo ratings to a 1500 ceiling.

Independent index
Artificial Analysis Intelligence Index
Composite of ~9–10 independent evals · Artificial Analysis
16/ 100
BenchmarkResult
Artificial Analysis Intelligence Indexindependent
GeneralArtificial Analysis
16%
τ-benchindependent
Agenticτ-bench
70.2%
LMArena Eloindependent
GeneralLMArena
1418Elo
Alternatives

Comparable models

Models in a roughly similar class — worth weighing against this one.

Llama 4 MaverickMeta AI

Open natively-multimodal MoE flagship — a strong, cheap general workhorse.

DeepSeek-V3.2DeepSeek

Ultra-cheap open MoE with standout math and efficiency.

Qwen3.7-MaxAlibaba Qwen

Alibaba's proprietary agentic reasoning flagship with a 1M context.