Xiaomi: MiMo-V2-Flash

Generates text using a 309B-parameter Mixture-of-Experts architecture that activates only 15B parameters per token, routing inputs through learned gating networks to specialized expert sub-models. This sparse activation pattern reduces computational cost during inference while maintaining model capacity through conditional expert selection, enabling efficient token generation for long-context conversations and multi-turn dialogue without full model computation.

Processes input sequences using a hybrid attention architecture that combines local (windowed) attention for nearby tokens with sparse global attention for distant dependencies, reducing quadratic attention complexity to near-linear while preserving long-range semantic relationships. This pattern enables efficient processing of longer contexts than standard dense attention while maintaining coherence across document-length inputs.

Generates coherent text across multiple languages (Chinese, English, and others) using a unified tokenizer and shared embedding space, enabling code-switching and cross-lingual reasoning without language-specific model branches. The model learns language-agnostic representations that allow seamless transitions between languages within a single generation pass.

Delivers generated text incrementally via HTTP streaming endpoints (compatible with OpenRouter), returning tokens as they are produced rather than waiting for full completion. This pattern enables real-time display of model output, reduces perceived latency in user-facing applications, and allows clients to interrupt generation early if needed.

Processes multiple prompts or requests in batches through the OpenRouter API, amortizing overhead costs and potentially receiving volume-based pricing discounts. Batch processing groups requests together for efficient GPU utilization and reduced per-token costs compared to individual request handling.

Maintains and processes multi-turn conversation history to generate contextually appropriate responses that reference previous exchanges, user preferences, and established context. The model uses attention mechanisms to weight relevant historical context and avoid repetition or contradiction with earlier statements in the conversation.

Accepts system prompts and instruction-based conditioning to guide response generation toward specific styles, formats, or behaviors. The model uses the system prompt as a high-priority context that influences token generation throughout the response, enabling role-playing, format specification, and behavioral constraints without fine-tuning.

Generates text that conforms to specified JSON schemas or structured formats through prompt-based guidance or constrained decoding, enabling reliable extraction of structured data from unstructured inputs. The model uses schema information to bias token generation toward valid outputs that match the specified structure.