Mistral: Ministral 3 3B 2512

Generates coherent text responses to prompts while maintaining the ability to process and understand image inputs, using a 3B parameter architecture optimized for inference speed and memory efficiency. The model uses a transformer-based decoder with vision encoder integration that allows it to analyze images and incorporate visual context into text generation without requiring separate vision-language alignment layers typical of larger models.

Executes model inference through OpenRouter's REST API endpoints with support for token-by-token streaming responses, allowing real-time text generation without waiting for full completion. The implementation uses HTTP POST requests with JSON payloads and optional Server-Sent Events (SSE) streaming, enabling progressive output rendering in client applications and reduced perceived latency.

Processes images alongside text prompts to extract visual context and incorporate it into response generation, using an integrated vision encoder that converts image pixels into embedding space compatible with the language model's token representations. The model can reason about image content, answer questions about visual elements, and generate text that references specific details from provided images.

Maintains multi-turn conversation state by accepting arrays of message objects with role-based formatting (system, user, assistant), allowing the model to reference previous exchanges and maintain conversational coherence across multiple requests. The implementation uses a standard chat completion message format where each turn is encoded as a separate token sequence, with the model attending to all prior messages within its context window.

Exposes inference parameters (temperature, top_p, top_k, max_tokens) that control the randomness and length of generated text, allowing developers to tune output behavior from deterministic (temperature=0) to highly creative (temperature=2.0). The implementation uses standard sampling techniques where temperature scales logit distributions before softmax, and top_p/top_k apply nucleus and k-sampling filters to the token probability distribution.

Executes inference through OpenRouter's pricing model which charges separately for input and output tokens, with published rates visible before API calls. The model's 3B parameter size results in lower per-token costs compared to larger models, and OpenRouter's aggregation model allows price comparison across providers without switching infrastructure.