Alibaba’s artificial intelligence development team unveiled the open-source, multimodal AI model named Qwen3.6-35B-A3B on Wednesday. This launch reinforces the company’s commitment to strengthening its position in the advanced technology market and delivering innovative solutions for developers and enterprises that rely on artificial intelligence in their daily operations.
An innovative technical architecture for resource conservation
The new model relies on a technical architecture known as the sparse mixture of experts, a revolutionary technique that enables powerful and independent programming performance at a fraction of the computational cost required by larger dense models. This advanced model consists of 35 billion total parameters, but its standout feature is that it activates only 3 billion parameters during each inference or data processing operation.
This intelligent model design makes it much more efficient to operate compared to dense models offering similar performance, significantly reducing computational resource and energy consumption. Despite this economically and environmentally lightweight footprint, tests have demonstrated the model’s ability to handle the most complex programming scenarios requested by developers, marking a qualitative leap in the world of automated programming.
Superiority in benchmark evaluations
The new model achieved impressive results in benchmark tests, outperforming both its predecessor Qwen3.5-35B-A3B and competitor Google’s Gemma4-31B model in complex coding tasks. This dominance is evident in the following results:
- In Terminal-Bench 2.0, a widely used metric for evaluating autonomous AI coding capabilities, the model scored 51.5 points, leading Google’s model, which achieved 42.9 points.
- In the advanced SWE-bench Pro test, the model reached 49.5 points compared to approximately 35.7 points for Google’s model.
- In the SWE-bench Verified test, the model scored 73.4 points versus 52.0 points for the direct competitor.
Additionally, the model achieved top scores in other important evaluations such as QwenClawBench, NL2Repo, and Claw-Eval, proving its comprehensive superiority in its category and its exceptional ability to write and analyze code with unmatched efficiency.
Advanced reasoning features and improvements
These results represent a significant qualitative leap compared to the previous version released last February, which scored 67.0 points in documented evaluations. Alibaba’s development team highlighted the fundamental improvements made to frontend workflows and reasoning capabilities across entire code repositories.
Among the most notable new additions to the model is the “thought preservation” feature, which retains inference context across multiple conversation rounds. This innovative feature significantly reduces the computational burden and time spent during iterative software development, giving developers a smoother and more efficient experience that allows the AI to remember previous steps and build upon them without needing to reprocess the entire context from scratch.
Flexible options for enterprises and developers
Operationally and oriented toward enterprise support, the model has been released under the open Apache 2.0 license, making it available for commercial and research use with high flexibility. Developers can easily access the model via popular platforms such as Hugging Face, through the Qwen Studio platform, or via the official API.
To ensure a smooth transition for existing users, the new model shares the same core technical architecture as its predecessor, meaning existing tools used in open development environments, such as the vLLM tool, are expected to work with it directly without requiring any additional code modifications that might disrupt workflows.
This powerful release positions Qwen3.6-35B-A3B as a cost-effective option for enterprise developers looking to build AI-driven coding agents and develop multi-step workflows. The real competitive advantage lies in the model’s ability to run smoothly on consumer-grade GPUs when using quantization techniques, while maintaining its ability to compete with larger, more complex models. Finally, this model is noted to be the first open-weights model of the new release family, with additional sizes expected to be launched soon to meet the diverse needs of the growing tech market.
Frequently Asked Questions
What distinguishes the Qwen3.6-35B-A3B model from other models?
The model is distinguished by its use of the sparse mixture of experts technique, possessing 35 billion total parameters while consuming only 3 billion parameters per processing operation, delivering superior programming performance at a very low computational cost compared to other models.
How does the new model outperform Google’s model?
The model achieved notable superiority in globally adopted standard programming tests, scoring 51.5 points in the Terminal-Bench 2.0 test compared to 42.9 points for the Gemma4-31B model, and also outperformed it in other complex code-writing tests such as SWE-bench Pro.
What is the “thought preservation” feature added to the modern model?
It is a new technical feature that allows the model to retain the context of inference and logic used during previous conversation rounds, reducing computational overhead and accelerating iterative software development without reprocessing data from the start.
Can the model run on standard computers?
Yes, thanks to its high efficiency and lightweight operating size, the model can run on consumer-grade GPUs when using quantization techniques, making it ideal for startups and individual developers.
