A Chinese AI called 'Qwen3.6-35B-A3B,' which is more powerful than Gemma4, has been released as an open model.



Qwen (Tongyi Lab), the AI research team of the Chinese AI company Alibaba, announced '

Qwen3.6-Plus ' on April 2, 2026, and released the first open model in the series, ' Qwen3.6-35B-A3B, ' on April 15, 2026. Qwen3.6-35B-A3B is a highly efficient model that exhibits superior agent coding performance, and is touted as being more powerful than Google's open model suite, Gemma 4 .

Qwen3.6-35B-A3B: Agentic Coding Power, Now Open to All
https://qwen.ai/blog?id=qwen3.6-35b-a3b

Qwen3.6-35B-A3B on my laptop drew me a better pelican than Claude Opus 4.7
https://simonwillison.net/2026/Apr/16/qwen-beats-opus/







Qwen3.6-35B-A3B is an open-weight MoE model with a total of 35 billion parameters and only 3 billion active parameters, making it a highly efficient model, yet it outperforms Gemma4-31B, which loads 31 billion parameters at once.

The following are benchmarks published by Qwen, showing that it outperforms Gemma4-26B-A4B and Gemma4-31B in all 12 benchmarks.



According to Qwen, the Qwen3.6 series is designed with multimodal capabilities in mind from the start, and the Qwen3.6-35B-A3B exhibits perceptual and multimodal reasoning capabilities far beyond what one might expect from its scale of only about 3 billion active parameters. In most of the visual language benchmarks conducted by Qwen, the performance of the Qwen3.6-35B-A3B was reported to be equal to or better than that of Claude Sonnet 4.5. Its strengths are particularly evident in its ability to understand and describe spatial structures.

Furthermore, in a benchmark devised by engineer Simon Willison, which involved

drawing a pelican riding a bicycle , Willison reported that the Qwen3.6-35B-A3B produced a better pelican than the Claude Opus 4.7 . Below, the image on the left is from the Qwen3.6-35B-A3B, and the image on the right is from the Claude Opus 4.7; the bicycle in the Claude Opus 4.7 appears to have a distorted frame.



Wilson stated, 'The Pelican benchmark was used as a joke, but there was a strange correlation between the quality of the Pelicans produced and the general usefulness of the models. However, it's unlikely that the lighter Qwen3.6-35B-A3B is more powerful and practical than the latest model, the Claude Opus 4.7, which Anthropic released on April 17, 2026. This benchmark simply shows that 'if you want an SVG illustration of a Pelican riding a bicycle, the Qwen3.6-35B-A3B, which runs on a laptop, is better than the Claude Opus 4.7.''

Qwen3.6-35B-A3B features excellent agent coding capabilities and seamlessly integrates with popular third-party coding assistants such as OpenClaw, Claude Code, and Qwen Code. Instructions and scripts for connecting Qwen3.6-35B-A3B with OpenClaw are available on the official blog .

Qwen3.6-35B-A3B is available as an open wait for self-hosting with Hugging Face and ModelScope via the Alibaba Cloud Model Studio API . It can also be tried from Qwen Studio .

Qwen stated, 'We will continue to expand the Qwen 3.6 open-source family and push the boundaries of what can be achieved with efficient and open models. We appreciate the feedback from the community and look forward to seeing what you build with Qwen 3.6-35B-A3B. Stay tuned for future releases!'

Related Posts:

in AI, Posted by log1e_dh