Ministral 3 8B Instruct 2512

Architecture overview:
• Dense 8.4B parameter language backbone paired with a 0.4B vision encoder for unified text and image IO.
• 256K token context window shared with the rest of the Ministral 3 family for consistent long-context behavior.
• Instruction-tuned head optimized for assistants, agents, and structured outputs such as JSON and tool calls.

Training and performance:
• Trained on diverse multilingual, code, and web-style corpora to provide robust coverage in the 8B tier.
• Instruction tuning emphasizes helpfulness, harmlessness, and adherence to system prompts over raw perplexity.
• Positioned as a mid-size workhorse that delivers near-frontier quality for many assistant and analytic tasks at lower cost and latency.

Assistants and agents:
• General-purpose chat assistants for support, operations, and knowledge work where responsiveness and quality must balance cost.
• Multimodal internal copilots that combine screenshots, documents, and text queries for debugging, analysis, and investigation.
• Agentic systems that plan, call tools, and synthesize results into natural-language recommendations or summaries.

Product and platform use cases:
• Embedded chat and help widgets inside SaaS products and dashboards.
• RAG-style knowledge interfaces over product docs, knowledge bases, and semi-structured data using the 256K context.
• Content generation, rewriting, translation, and summarization workflows that need solid multilingual quality.