N
NexusRoute
Back to Models

Mistral Large 3

Mistralfrontier

Mistral's most capable model. A 675B MoE (41B active) with 256K context, native multimodal support, and full Apache 2.0 open-weight licensing. The strongest truly open-source frontier model. Excels at multilingual tasks with particular strength in European languages.

Released 2025-12-03Knowledge cutoff: 2025-09
Medium confidence|Updated 72d ago|85% source confidence

Specifications

Context Window

256K tokens

Max Output

32K tokens

Input Price

$2.00 / 1M tokens

Output Price

$6.00 / 1M tokens

Latency Tier

Moderate (speed score: 6/10)

Capability Profile

Reasoning
8.5/10
Coding
8.5/10
Long Context
8/10
Structured Output
8/10
Factuality
8/10
Instruction Following
8/10
Tool Use
8/10
Conversational
8/10
Creativity
7.5/10
Safety & Enterprise
7.5/10
Multimodal
7/10
Cost Efficiency
6.5/10
Speed
6/10

Feature Support

Vision Yes
Audio In No
Audio Out No
Video No
Image Generation No
Image Editing No
Function Calling Yes
JSON Mode Yes
Structured Output Yes
Streaming Yes
Reasoning Yes
Realtime No
Computer Use No
Web Search No

Best Use Cases

European enterprise deployments requiring open-source licensing and EU data residency
Multilingual applications spanning 20+ languages, especially European ones
Self-hosted deployments needing frontier quality with Apache 2.0 licensing
Function calling and agentic workflows with strong tool use
Code generation with multilingual documentation and comments

Not Ideal For

The absolute hardest coding tasks (Claude Opus is better)
Tasks requiring the longest context windows (Gemini/Grok have larger)
Audio or video processing
Ultra-budget workloads where DeepSeek is 7x cheaper

Strengths

Apache 2.0 license — the most permissive license among frontier-class models
Excellent multilingual support, especially French, German, Spanish, Italian, Portuguese
Strong function calling and agentic tool use
EU data residency through Mistral's La Plateforme
Good all-around quality competitive with GPT-5.4-mini tier
Vision support adds multimodal capability to the open-source frontier

Weaknesses

Coding quality strong but below Claude Opus/Sonnet 4.6 on SWE-bench
256K context is large but smaller than competitors offering 1M+
Self-hosting requires significant GPU infrastructure (8+ H100s)
Less widely adopted than Llama or GPT — smaller ecosystem
English-only tasks may prefer GPT or Claude for slightly better quality

Edge Cases & Notes

Apache 2.0 licensing makes it unique among frontier models for commercial use without restrictions
41B active parameters per token despite 675B total — MoE efficiency
Particularly strong for French-language tasks — Mistral's home-language advantage
EU data residency is a genuine differentiator for European enterprises

Provider Notes

Available through Mistral's La Plateforme, Azure AI, and self-hosted. Apache 2.0 license allows unrestricted commercial use. EU data residency available on La Plateforme. Best choice for European enterprise with open-source requirements.

Benchmarks

MMLU88%
HumanEval87.5%
Arena Elo1330

Benchmark Notes

MMLU-Pro 88%. Strong multilingual benchmarks, especially on European languages. SWE-bench ~46%. Competitive with Llama 4 Maverick on most benchmarks. Apache 2.0 makes it unique in its quality tier.

Research Meta

Last Evaluated

2026-03-15

Source Confidence

85%

Evaluation Method

LMSYS Arena, MMLU-Pro, multilingual evaluations, SWE-bench, open-source license comparison

Needs Re-evaluation

No

Sources

  • Mistral Large 3 technical report
  • LMSYS Chatbot Arena
  • Open LLM Leaderboard
  • Mistral AI benchmarks

Continue exploring

Route a prompt

See how Mistral Large 3 ranks

Compare models

Side-by-side analysis

Browse registry

Explore all 24 models