Alibaba Qwen3.6 Beats Google Gemma 4 on Coding Benchmarks

Alibaba launched Qwen3.6-35B-A3B, a fresh open AI model. This mixture-of-experts design activates only three out of 35 billion parameters during operation. That approach lowers compute requirements while maintaining performance levels, Alibaba reports.

The company states this version surpasses its prior model, Qwen3.5-35B-A3B, in agentic coding tasks. Compared to Google's open Gemma 4-31B, Qwen3.6-35B-A3B takes the lead in all listed coding evaluations. Scores reach 73.4 against 52.0 on SWE-bench Verified. On Terminal-Bench 2.0, results stand at 51.5 versus 42.9.

Strong Results in Reasoning and More

Qwen3.6-35B-A3B also shows gains in reasoning benchmarks. It scores 86.0 on GPQA, ahead of Gemma 4-31B's 84.3. For AIME26, the model hits 92.7 compared to 89.2. Alibaba notes it holds its own against Claude Sonnet 4.5 in image and video processing.

Charts from benchmarks confirm Qwen3.6-35B-A3B's advantages over Qwen3.5-35B-A3B and Google's Gemma 4 models in coding, reasoning, and multimodal categories.

Alibaba, founded in 1999 as a major Chinese e-commerce firm, has expanded into cloud computing and AI research through its DAMO Academy. The Qwen series represents its push into open-source large language models, starting with Qwen1.5 and building through versions like Qwen2 and Qwen3. These models compete in a field crowded with efforts from companies worldwide aiming to provide accessible AI tools.

Google's Gemma line, introduced as lightweight open models, seeks to offer developers efficient options. Gemma builds on earlier releases to support various applications without heavy resource demands.

Features and Availability

Users find both thinking and non-thinking modes in the model. They can test it in Qwen Studio. API access comes as Qwen3.6 Flash via Alibaba Cloud Model Studio. Weights download from Hugging Face and ModelScope. This release comes after the bigger Qwen3.6-Plus.

Agentic coding benchmarks test AI abilities in real software engineering scenarios. SWE-bench Verified evaluates fixes for GitHub issues. Terminal-Bench 2.0 checks command-line tasks. GPQA involves graduate-level questions in physics, biology, and chemistry. AIME26 draws from math competition problems.

Mixture-of-experts architecture, common in modern AI, routes inputs to specialized sub-networks. Activating few experts saves power and speeds inference, key for practical use.

Alibaba positions Qwen3.6-35B-A3B as a strong contender among open models. Its broad benchmark wins highlight progress in efficient, high-performing AI.

Background on Key Players

Hugging Face hosts thousands of open models, serving as a central repository for AI developers. ModelScope, Alibaba's platform, focuses on Chinese-language support alongside global models. Alibaba Cloud provides enterprise AI services, competing with leaders like AWS and Azure.

This development fits into ongoing races for AI supremacy, where open models gain traction for transparency and customization.

Related on Neura Market

• AI Models Directory
• Open Source LLMs
• AI Benchmarks Hub