Alibaba Releases Qwen3.5 Model for Agentic AI Era

Alibaba unveiled the Qwen3.5 model on February 16 2026 through its Qwen team on its official blog. The release marks the debut of the first open-weight model in the Qwen3.5 series, known as Qwen3.5-397B-A17B.

Credit: Qwen

This native vision-language model delivers top results across benchmark evaluations in reasoning coding agent capabilities and multimodal understanding.

The Qwen team stated in its official blog post We are delighted to announce the official release of Qwen3.5 introducing the open-weight of the first model in the Qwen3.5 series namely Qwen3.5-397B-A17B.

The new Qwen AI model builds on a hybrid architecture that combines linear attention through Gated Delta Networks with a sparse mixture-of-experts structure.

It features 397 billion total parameters but activates just 17 billion during each forward pass which boosts inference speed and lowers costs. The model now supports 201 languages and dialects up from 119 in prior versions.

Qwen3.5 targets the agentic AI era where systems handle complex tasks autonomously. The model includes visual agentic capabilities that let it interact directly with smartphones and computers.

• On mobile devices it follows natural-language instructions to perform actions inside apps and switch between them seamlessly.
• On desktops it manages extended workflows for office automation. It processes inputs up to one million tokens which covers about two hours of video.

This enables applications such as converting hand-drawn user interface sketches into frontend code reverse-engineering game logic from footage and summarizing long videos into structured web pages.

The architecture draws from the Qwen3-Next foundation with higher-sparsity mixture-of-experts gated attention mechanisms stability optimizations and multi-token prediction.Under 32k context length the decoding throughput reaches 8.6 times that of Qwen3-Max and 19 times under 256k context.

Compared to the Qwen3-235B-A22B model the new version achieves 3.5 times the throughput at 32k and 7.2 times at 256k.

Benchmark results show the model leading in multiple categories.

Credit: Qwen

In language tasks it scores 87.8 on MMLU-Pro 94.9 on MMLU-Redux 70.4 on SuperGPQA 93.0 on C-Eval 92.6 on IFEval and 83.6 on LiveCodeBench v6.

For vision-language evaluations it posts 85.0 on MMMU 79.0 on MMMU-Pro 90.3 on Mathvista and 93.7 on MMBench EN-DEV-v1.1.

Agent-specific tests include 76.4 on SWE-bench Verified 72.9 on BFCL-V4 and 65.6 on ScreenSpot Pro.

The Qwen3.5-Plus hosted version became available immediately on Alibaba Cloud Model Studio with a one million token context window and built-in tools for adaptive use.

The open-weight Qwen3.5-397B-A17B is downloadable on platforms including Hugging Face ModelScope and GitHub.

Users can access it through the Qwen Chat app in auto thinking or fast modes.

Qwen3.5 provides a strong foundation for universal digital agents through its efficient hybrid architecture and native multimodal reasoning.

The model can use tools such as a code interpreter and image search during reasoning to interpret visuals transform images render intermediate steps and verify results with external data.