Alibaba Open-Sources Qwen3.6-35B-A3B

Alibaba's Qwen team has just open-sourced Qwen3.6-35B-A3B, the first open-weight model from the Qwen3.6 series. This model features a total of 35 billion parameters but activates only about 3 billion per token, leveraging a Mixture of Experts (MoE) architecture for efficiency. It significantly surpasses its predecessor, Qwen3.5-35B-A3B, in agent coding tasks, with Terminal-Bench 2.0 scores jumping from 40.5 to 51.5 and QwenWebBench Elo ratings rising from 978 to 1397, approaching or exceeding the previous dense flagship, Qwen3.5-27B. Released under the Apache 2.0 license, the model is now available on HuggingFace and ModelScope, with an API version hosted as qwen3.6-flash.

Background: First Open-Source Release Since Key Departures

The Qwen team has experienced recent personnel changes, with core developer Lin Junyang and several other executives leaving Alibaba earlier in 2026. Prior to this, Qwen3.6-Plus was launched as an API-only model on OpenRouter in March, setting records for usage but without releasing weights. Today's open-sourcing of Qwen3.6-35B-A3B marks the first accessible model from the Qwen3.6 lineup.

Core Features: Efficiency Gains Without Sacrificing Performance

MoE models typically trade off some capability for inference efficiency compared to dense models of similar total parameters. However, Qwen3.6-35B-A3B shows substantial improvements, especially in agent coding, rivaling or surpassing Qwen3.5-27B in several benchmarks. For instance, on Terminal-Bench 2.0, it scores 51.5 versus Qwen3.5-27B's 41.6, though it slightly trails on SWE-bench Verified (73.4 vs 75.0). Compared to its direct predecessor, Qwen3.5-35B-A3B, gains are most pronounced in agent execution tasks.

Agent Coding and Language Ability Evaluations

Progress is concentrated in agent coding, with steady improvements in knowledge and reasoning, while general agent performance is mixed.

Agent Coding

This is the standout area, with Qwen3.6-35B-A3B excelling across nearly all coding agent benchmarks. It outperforms Qwen3.5-35B-A3B and approaches Qwen3.5-27B levels. Key scores include:

• SWE-bench Verified: 73.4 (vs 75.0 for Qwen3.5-27B and 70.0 for predecessor)
• Terminal-Bench 2.0: 51.5 (vs 41.6 for Qwen3.5-27B and 42.9 for Gemma4-31B)
• QwenWebBench Elo: 1397 (from 978 previously, indicating major improvements in generating runnable frontend code)

General Agent

Performance in general agent tasks like customer service and tool calling is uneven. For example, TAU3-Bench scores dipped slightly to 67.2 from 68.9, but MCPMark improved to 37.0 from 27.0, showing better accuracy in MCP tool invocation.

Knowledge and Reasoning

Steady enhancements are seen in knowledge and reasoning benchmarks. GPQA Diamond rose to 86.0 from 84.2, surpassing Qwen3.5-27B's 85.5, and LiveCodeBench v6 increased to 80.4 from 74.6, nearly matching the dense model's 80.7. Chinese proficiency remains stable with C-Eval at 90.0.

Vision-Language Capabilities: Outperforming Claude Sonnet 4.5 in Multiple Areas

As a natively multimodal model, Qwen3.6-35B-A3B shows strong visual and document understanding. It surpasses Claude Sonnet 4.5 in benchmarks like RealWorldQA (85.3 vs 70.3) and OmniDocBench1.5 (89.9 vs 85.8). Gains are also evident over its predecessor, with improvements in spatial intelligence and video understanding, such as RefCOCO scores reaching 92.0 from 89.2 and ODInW13 jumping to 50.8 from 42.6.

How to Obtain and Deploy

The model is fully open-sourced and available for download on HuggingFace (Qwen/Qwen3.6-35B-A3B) and ModelScope under the Apache 2.0 license, allowing unrestricted commercial use. An API version named qwen3.6-flash is hosted on Alibaba Cloud, compatible with both OpenAI and Anthropic protocols, enabling integration with tools like Claude Code. It also supports agent coding frameworks such as OpenClaw and Qwen Code, and can be tested on Qwen Studio. For more detailed benchmarks and comparisons, refer to DataLearner's model information card.