Qwen3.6-35B-A3B: A model with agentic coding performance is released to everyone

Alibaba's Qwen team has released Qwen3.6-35B-A3B, an open-source model based on the MoE (Mixture-of-Experts) architecture. This model has a total of 35 billion parameters, but activates only 3 billion (3B) parameters during actual inference, providing high efficiency.

Agentic Coding and Multimodal Performance

• Agentic Coding: The ability to diagnose and fix its own code has been strengthened, surpassing the existing larger Dense model Qwen3.5-27B on major benchmarks such as SWE-bench Verified (73.4 points).
• Multimodal: It processes text, images, and video in an integrated way, and showed performance surpassing Claude Sonnet 4.5 in MMMU (81.7 points) and spatial intelligence (RefCOCO 92.0).
• Math and Coding: It recorded results on par with large Dense models on AIME 2026 (92.7 points) and LiveCodeBench v6 (80.4 points).

Key Strengths and Differentiators

• Resource Efficiency: With only 3B active parameters, it maintains high performance while consuming less GPU memory and power.
• Ecosystem Compatibility: It supports the Anthropic API protocol, allowing immediate integration with third-party coding tools such as Claude Code.
• Context Retention: Through the preserve_thinking feature, it preserves the thinking process from previous conversations during agent tasks, which is advantageous for maintaining long-term context.

Limitations

• Performance is lower than large Dense models on general-purpose agent tasks (VITA-Bench) and highly difficult academic reasoning (HLE).
• API service is currently in a pre-launch state.