Qwen: Qwen3.5 Plus 2026-02-15

Processes images, text, and video inputs simultaneously using a hybrid architecture combining linear attention mechanisms with sparse mixture-of-experts routing. Linear attention reduces computational complexity from O(n²) to O(n) while sparse MoE selectively activates expert parameters based on input type and content, enabling efficient processing of high-resolution visual content alongside text without full model activation.

Processes video inputs by decomposing them into frame sequences and applying vision-language understanding across temporal boundaries. The sparse MoE architecture selectively activates video-specialized experts when video tokens are detected, enabling efficient analysis of motion, scene changes, and temporal relationships without processing every frame through the full model capacity.

Implements sparse mixture-of-experts routing that dynamically selects which expert parameters activate based on input content type and complexity, reducing per-token computation from full model capacity to a fraction of parameters. The routing mechanism uses learned gating functions to assign tokens to specialized experts (vision, language, multimodal), enabling high-throughput inference without loading all parameters for every request.

Processes high-resolution images using linear attention mechanisms that scale O(n) instead of O(n²), enabling efficient encoding of dense visual tokens without memory explosion. The architecture decomposes image patches into token sequences and applies linear attention transformations, allowing processing of images with significantly more pixels than quadratic-attention models while maintaining spatial reasoning capability.

Generates and understands text across multiple languages using a shared token vocabulary and language-agnostic attention mechanisms. The model applies the same linear attention and sparse MoE routing to all languages, with language-specific expert routing enabling efficient multilingual inference without separate model instances per language.

Extracts structured information (JSON, tables, key-value pairs) from unstructured text and images using prompt-based schema specification and constrained decoding. The model applies vision-language understanding to identify relevant content regions, then generates structured output conforming to specified schemas, with optional validation against provided JSON schemas.

Analyzes and generates code across multiple programming languages using vision-language understanding to parse code syntax from images and text, combined with language-specific expert routing in the MoE layer. Supports code completion, explanation, and refactoring by maintaining semantic understanding of code structure and applying language-specific reasoning patterns.

Performs multi-step reasoning and problem decomposition using chain-of-thought patterns and planning-aware expert routing. The sparse MoE architecture activates reasoning-specialized experts when processing complex queries, enabling step-by-step problem solving with explicit intermediate reasoning steps that improve accuracy on tasks requiring logical inference.

Provides HTTP/REST API access to the model with support for both streaming (token-by-token) and batch inference modes. Streaming responses enable real-time output display and early termination, while batch mode optimizes throughput for non-latency-sensitive workloads. The API abstracts underlying sparse MoE routing and linear attention mechanisms, exposing a standard interface compatible with OpenAI API conventions.