Qwen3.5: Toward Native Multimodal Agents

Qwen3.5-397B-A17B has been released as the first open-weight model in the Qwen3.5 series. As a native vision-language model, it delivers strong performance across reasoning, coding, agentic capabilities, and multimodal understanding overall, and it also improves efficiency through a structure in which only 17B of its total 397B parameters are activated.

The core architecture is a hybrid design combining Gated Delta Networks-based linear attention with sparse MoE. This reduces speed and cost while suppressing performance degradation, and it also expands language and dialect support from 119 to 201.

The performance table compares a wide range of benchmarks, including language, STEM, reasoning, general agent, search agent, multilingual, coding agent, vision language, video understanding, visual agent, and medical VQA. The model recorded top-tier results on several items, and the improvements are especially emphasized in multimodal tasks and agentic-type work.

Pretraining is organized around three axes.

• Power: Trained on a much larger scale of visual-text tokens compared to Qwen3, allowing Qwen3.5-397B-A17B to reach a level comparable to Qwen3-Max-Base, which is in the 1T+ parameter class.
• Efficiency: Built on the Qwen3-Next architecture, applying a higher-sparsity MoE, a Gated DeltaNet + Gated Attention hybrid, stabilization optimizations, and multi-token prediction.
• Versatility: Implements native multimodality through early text-vision fusion, expands visual/STEM/video data, and improves encoding/decoding efficiency with a 250k vocab.

Inference efficiency has also been greatly improved. At 32k/256k context, the decoding throughput of Qwen3.5-397B-A17B is reported to be 8.6x/19.0x that of Qwen3-Max, respectively, and 3.5x/7.2x that of Qwen3-235B-A22B.

On the reinforcement learning side, the model expanded RL tasks and environments as broadly as possible to raise general agent capability. It also introduced an asynchronous RL framework and a fully decoupled training-inference structure to improve hardware utilization, load balancing, and failure recovery, claiming an overall 3x-5x end-to-end speedup.

On the product side, Qwen Chat offers three modes: auto, thinking, and fast. Qwen3.5-Plus is available on Alibaba Cloud Model Studio, where enable_thinking and enable_search can be used to enable reasoning, web search, and Code Interpreter features.