HeadsUpAI

Alibaba Qwen Launches Qwen3.6-Max-Preview With Enhanced Agentic Coding Reliability

· Updated

Qwen, the AI lab behind the Qwen series of models, released an early preview of Qwen3.6-Max-Preview, the upcoming flagship in its latest series. This proprietary release follows the production-ready Qwen3.6-Plus and focuses on agentic coding (AI that autonomously writes and iterates on code).
SkillsBench gain
+9.9 points
SciCode gain
+6.3 points
NL2Repo gain
+5.0 points
Terminal-Bench 2.0 gain
+3.8 points
SuperGPQA gain
+2.3 points
API parameter
preserve_thinking
Availability
Qwen Studio, API (coming soon)

The update addresses a critical bottleneck in agentic reliability: the loss of logic during long-running tasks. By introducing a preserve_thinking parameter, the model maintains its internal reasoning traces across an entire conversation. This builds on the agentic reliability focus established in earlier previews, ensuring the model doesn't lose context during multi-step edits.

You can test the model immediately via the interactive chat interface in Qwen Studio. Developers should look for the qwen3.6-max-preview identifier in the Alibaba Cloud Model Studio API, which is coming soon. The API supports industry-standard protocols and includes specific configurations for enabling and preserving thinking tokens during inference.

View the full update on qwen.ai
Qwen
Qwen
@Alibaba_Qwen
X

🚀 Introducing Qwen3.6-Max-Preview, an early preview of our next flagship model Highlights： ⚡️ Improved agentic coding capability over Qwen3.6-Plus 📖 Stronger world knowledge and instruction following 🌍 Improved real-world agent and knowledge reliability performance Smarter, sharper, still evolving. More Qwen3.6 models to come. Stay tuned! 🔗👇 Blog: https://t.co/6hDQJhmkjM Qwen Studio: https://t.co/Fe2X1IrW6r API: https://t.co/xWPs39LBIm

308retweets3.1klikes
View on X

Still wondering? A few quick answers below.

Qwen3.6-Max-Preview is an early preview of Alibaba's next flagship proprietary model. It is designed to be the most capable model in the Qwen3.6 series, offering significant improvements in world knowledge, instruction following, and agentic coding. As a preview release, it is still under active development with further performance gains expected in future iterations.

Compared to the production-ready Qwen3.6-Plus, the Max-Preview version delivers substantial performance gains across several benchmarks. It shows a 9.9-point improvement on SkillsBench and a 6.3-point increase on SciCode. These updates specifically target agentic coding reliability and world knowledge, making it more effective for complex, multi-step tasks that require deep reasoning and repository-level understanding.

The preserve_thinking feature allows the model to maintain its internal reasoning traces, or thinking tokens, from all preceding turns in a conversation. Alibaba recommends this setting for agentic tasks because it helps the model stay consistent during long-running workflows. By preserving these traces, the model can better track its logic and goals across complex, multi-turn interactions.

You can currently use Qwen3.6-Max-Preview for interactive chat through the Qwen Studio web interface. For developers, the model will soon be available via the Alibaba Cloud Model Studio API using the model identifier qwen3.6-max-preview. The API is compatible with industry-standard specifications from OpenAI and Anthropic, allowing for easy integration into existing AI agent frameworks.

No, Qwen3.6-Max-Preview is a proprietary model hosted by Alibaba Cloud. While the Qwen team frequently releases open-weight models like the Qwen3.6-35B-A3B, the Max-Preview version is currently restricted to the company's hosted platforms. It is available as a managed service through Qwen Studio and the Alibaba Cloud Model Studio API rather than as a downloadable model file.

Share this update