Mistral Large 3
Mistral · mistral family · Official docs
Mistral Large 3 is the European frontier contender and the first truly open-weight model at the 675B scale. Its 256K context window is a standout feature, tied with Claude for the largest commercially available context. The function calling and structured output support is native and well-tested — Mistral has invested heavily in agentic capabilities. For Refrase users in regulated European industries, Mistral offers a GDPR-friendly alternative to US and Chinese providers. The Apache 2.0 licensing is a strategic differentiator: this is the only 675B-class model you can legally self-host and modify. The anti-patterns guidance from Mistral's docs (avoid vague quantifiers, prefer worded scales) is unusually practical and directly applicable to Refrase prompt optimization.
Specifications
Strengths
Key capabilities
- ✓Granular Mixture-of-Experts: 675B total parameters with 41B active (Language model: 673B/39B active, Vision encoder: 2.5B) (source: Hugging Face, Mistral-Large-3-675B-Instruct-2512 Model Card)
- ✓256K context window — one of the largest available, supporting extensive document processing and complex agentic workflows (source: Mistral Docs, Mistral Large 3 page)
- ✓Multimodal: native vision capabilities for image analysis with up to 10 images per prompt (source: Hugging Face, Mistral-Large-3-675B-Instruct-2512 Model Card)
- ✓Native function calling and structured JSON output — best-in-class agentic capabilities with tool-use support (source: Mistral Docs, Mistral Large 3 page)
- ✓Broad multilingual support: English, French, Spanish, German, Italian, Portuguese, Dutch, Chinese, Japanese, Korean, Arabic, and more (source: Mistral Docs, Mistral Large 3 page)
- ✓Open-weight Apache 2.0 license — fully open-sourced, marking Mistral's first frontier-class model under permissive license (source: Hugging Face, Mistral-Large-3-675B-Instruct-2512 Model Card)
- ✓Fill-in-the-Middle (FIM) for code generation, OCR with structured annotations, and audio transcription (source: Mistral Docs, Mistral Large 3 page)
- ✓Trained from scratch on ~3,000 H200 GPUs — frontier-level training compute (source: Hugging Face, Mistral-Large-3-675B-Instruct-2512 Model Card)
Known limitations
- ⚠Not a dedicated reasoning model — specialized reasoning models (like Magistral or Claude with extended thinking) may outperform on pure reasoning tasks (source: Hugging Face, Mistral-Large-3-675B-Instruct-2512 Model Card)
- ⚠Lags behind vision-first models on multimodal-optimized benchmarks despite supporting vision (source: Hugging Face, Mistral-Large-3-675B-Instruct-2512 Model Card)
- ⚠Extremely resource-intensive deployment: requires 8xH200 (FP8) or 8xH100/A100 (NVFP4) for inference (source: Hugging Face, Mistral-Large-3-675B-Instruct-2512 Model Card)
- ⚠Image aspect ratio sensitivity: best performance near 1:1 width-to-height ratio (source: Hugging Face, Mistral-Large-3-675B-Instruct-2512 Model Card)
How to prompt Mistral Large 3
Preferred instruction format
Standard chat format with system/user/assistant roles. System messages define environment and guidance. Mistral recommends defining roles explicitly: 'You are a [role], your task is to [task].' Supports Markdown and XML-style tags for structured instructions.
Recommended practices
- Use temperature < 0.1 for production/daily-driver use cases; higher temperature for creative tasks (source: Hugging Face, Mistral-Large-3-675B-Instruct-2512 Model Card)
- Define roles explicitly with 'You are a [role], your task is to [task]' framing (source: Mistral Docs, Prompting Capabilities)
- Organize instructions hierarchically with clear sections and subsections using Markdown or XML tags (source: Mistral Docs, Prompting Capabilities)
- Use few-shot prompting with example user-assistant exchanges for format-critical tasks (source: Mistral Docs, Prompting Capabilities)
- Keep tool definitions minimal — avoid overloading the model with too many function definitions (source: Hugging Face, Mistral-Large-3-675B-Instruct-2512 Model Card)
- Treat prompts like code: iterate, test, and evaluate changes systematically (source: Mistral Docs, Prompting Capabilities)
Anti-patterns to avoid
- Avoid vague quantifiers ('too long', 'many', 'few') — use objective measurements (source: Mistral Docs, Prompting Capabilities)
- Never use ambiguous descriptors like 'interesting' or 'better' without precise definitions (source: Mistral Docs, Prompting Capabilities)
- Do not use contradictory rules — use decision trees for conditional logic (source: Mistral Docs, Prompting Capabilities)
- Avoid asking the model to count characters or items — provide counts directly (source: Mistral Docs, Prompting Capabilities)
- Prefer worded scales ('Very Low', 'Good') over numeric scales (1-5) for evaluation tasks (source: Mistral Docs, Prompting Capabilities)