Mistral: Mistral Small 3.1 24B

Generates coherent, contextually-aware text responses to user prompts using a 24B parameter transformer architecture trained on instruction-following datasets. The model processes input tokens through multi-head attention layers and produces output via autoregressive decoding, optimized for chat and reasoning tasks through instruction-tuning on curated conversational and analytical datasets.

Processes both text and image inputs simultaneously to generate contextually-aware responses that reference visual content. The model integrates a vision encoder (likely CLIP-based or similar) that converts images into token embeddings, which are concatenated with text token embeddings and processed through the shared transformer backbone, enabling tasks like image captioning, visual question-answering, and scene understanding.

Exposes the model through OpenRouter's HTTP API with support for streaming token-by-token responses via Server-Sent Events (SSE) or chunked transfer encoding. Requests are routed through OpenRouter's load balancer to available Mistral Small 3.1 instances, with response streaming enabling real-time token delivery for interactive applications without waiting for full completion.

Maintains conversation history across multiple turns by accepting a messages array where each turn includes role (user/assistant/system) and content. The model processes the full conversation history as context, using attention mechanisms to weight recent messages more heavily while retaining earlier context, enabling coherent multi-turn dialogue without explicit memory management by the client.

Accepts hyperparameters (temperature, top_p, top_k, max_tokens, frequency_penalty, presence_penalty) that control the sampling strategy during token generation. Temperature scales logits before softmax to adjust randomness; top_p and top_k filter the token distribution; penalties discourage repetition. These parameters are applied during the autoregressive decoding loop, allowing fine-grained control over output diversity and length without model retraining.

Accepts optional JSON schema or format hints in system prompts to guide the model toward producing structured outputs (JSON, XML, YAML) that conform to specified schemas. The model uses instruction-tuning to recognize format requests and generate valid structured text, though without hard constraints—invalid JSON may still be produced if the model fails to follow the format instruction.