Alibaba's Qwen3.6-35B-A3B Outperforms Google Gemma 4 26B A4B in Agentic Coding Benchmarks
Qwen's new MoE model scores 73.4% on SWE-bench, crushing Gemma 4's 52%...
Alibaba's Qwen team has released Qwen3.6-35B-A3B, an open-weight Mixture-of-Experts (MoE) model that sets a new benchmark in agentic coding. In head-to-head testing on SWE-bench Verified, Qwen scored 73.4%, significantly outperforming Google's Gemma 4 26B A4B, which managed only 52.0%. This 21-point gap is striking because Qwen activates just 3 billion parameters per token versus Gemma's 4 billion, demonstrating superior architectural efficiency and real-world coding prowess.
The model's performance highlights Alibaba's growing lead in open-weight AI for developer tools and autonomous agents. With a leaner activation footprint, Qwen3.6-35B-A3B can handle complex software engineering tasks—like code generation, debugging, and refactoring—more accurately and cost-effectively. For enterprises building AI coding assistants or automated DevOps pipelines, this means better results with fewer computational resources, challenging Google's dominance in the agentic AI space.
- Qwen3.6-35B-A3B scored 73.4% on SWE-bench Verified, vs. Gemma 4's 52.0%
- Qwen activates only 3B parameters per token, compared to Gemma's 4B
- The model is open-weight, enabling customization for enterprise agentic coding tasks
Why It Matters
Open-weight coding models now beat Google's best, enabling cheaper, more capable AI agents for developers.