1. Official Launch of the Qwen3.6 Series

On April 2, 2026, Alibaba officially launched the new generation large language model Qwen3.6-Plus, the first model in the Qwen 3.6 series. Compared to the previous generation Qwen3.5, the new model shows significant overall performance improvements, especially achieving comprehensive leaps in programming capabilities, agent capabilities, and tool calling capabilities.

On the day after the launch (April 3), on the Code Arena leaderboard of the global LLM evaluation platform LMArena, Qwen3.6-Plus ranked second globally, becoming the highest-ranked Chinese large model on this leaderboard. More notably, within one day of launch (April 4), the model's daily call volume on the global LLM API platform OpenRouter exceeded 1.4 trillion tokens, topping the daily leaderboard and breaking the global record for single-model daily call volume on the platform.

2. Excellent Programming and Agent Capabilities

The most prominent feature of the Qwen3.6 series is its powerful programming capabilities. In authoritative evaluations of programming capabilities, Qwen3.6 surpassed models like GLM-5 and Kimi-K2.5 with 2 to 3 times the parameter count, becoming the strongest domestic programming model currently, approaching the global strongest programming models in the Claude series.

Specifically, in SWE-bench series real-world programming task tests, Terminal-Bench 2.0 examining terminal programming, long-horizon programming tasks NL2Repo, and real-world Agent capability evaluations such as Claw-Eval and QwenClawBench, Qwen3.6-Plus performance can fully rival Claude Opus 4.5. The model demonstrates excellent engineering deployment capabilities, not only capable of high-quality code repair but also skilled in complex terminal operations and automated task execution, achieving optimal results in multiple high-difficulty long-horizon planning tasks.

3. First Open Source Model: Qwen3.6-35B-A3B

Following the launch of Qwen3.6-Plus, the Alibaba Tongyi Qianwen team announced the open source of the Mixture of Experts (MoE) model Qwen3.6-35B-A3B on April 16. The model has a total parameter count of 35 billion, with only 3 billion activated parameters, focusing on lightweight efficiency and agent programming capabilities.

Model Architecture Features:

  • Adopts sparse MoE architecture, 35 billion total parameters, activating only 3 billion parameters per inference
  • Contains 40 layers, hidden dimension 2048, adopts mixed attention design
  • Native support for 262,144 token context, scalable up to 1,010,000 tokens
  • Supports multimodal thinking and non-thinking modes

4. Comprehensive Performance Superiority

In benchmark tests, Qwen3.6-35B-A3B achieved results close to or even surpassing 27 billion dense models with only 3 billion activated parameters:

Programming Capabilities:

  • SWE-bench Verified reached 73.4 (Qwen3.5-27B was 75.0)
  • Terminal-Bench 2.0 reached 51.5, surpassing all comparable peer models
  • Achieved 29.4 on NL2Repo, exceeding Qwen3.5-27B's 27.3

Vision-Language Capabilities:

  • Performance on most vision-language benchmarks is on par with Claude Sonnet 4.5, surpassing in some tasks
  • MMMU 81.7, MathVista 86.4, RealWorldQA 85.3, OmniDocBench 89.9
  • Advantages are particularly prominent in spatial intelligence: RefCOCO 92.0, ODInW13 50.8

5. Technical Innovation and Practical Features

The Qwen3.6 series is built based on community feedback, focusing on stability and actual development utility, with major upgrades concentrated in two aspects:

  1. Significantly Enhanced Agent Programming Capabilities: The model performs more smoothly and accurately on frontend workflows and repository-level reasoning tasks.

  2. New Thinking Retention Mechanism: Allows retaining historical reasoning context across multi-turn conversations, reducing repetitive overhead in iterative development.

The model has been adapted for third-party coding assistants such as OpenClaw, Qwen Code, and Claude Code, supporting integration via the Anthropic API protocol. The API adds a "preserve_thinking" function, which can retain all prior turn thinking content in messages; officially recommended for agent tasks.

6. Acquisition and Usage

Currently, the Qwen3.6 series is available through multiple channels:

  1. Qwen3.6-Plus: Available on Alibaba Cloud Bailian, with a minimum cost of 2 yuan per million input tokens; also launched on Wukong, Qwen App, and other Alibaba AI applications and platforms.

  2. Qwen3.6-35B-A3B: Weights are open source, supporting local deployment. Users can obtain it from the following platforms:

7. Industry Impact and Future Outlook

The launch of Qwen3.6 marks a new height for domestic large models in programming capabilities. IDC China Research Director Lu Yanxia stated that strengthening Agent capabilities is an important development direction for foundation large models in 2026, potentially including deep research, intelligent office, AI coding assistants, and other general scenarios.

Alibaba revealed that the Qwen 3.6 series will subsequently open source other model sizes, and the flagship model Qwen3.6-Max with stronger performance is also planned for release soon. With the open source of Qwen3.6, the global developer community will be able to more conveniently experience and utilize this advanced programming model, promoting application innovation in the field of AI software development.

The success of Qwen3.6 not only demonstrates China's rapid progress in large model technology but also provides efficient and practical open source options for the global AI community, expected to further promote the development and popularization of agent programming technology and applications.