Alibaba Launches Qwen3.7-Max Reasoning Agent with 1M Context Window

Alibaba has officially introduced Qwen3.7-Max, a proprietary reasoning agent model designed to handle complex, long-horizon tasks and multi-step autonomous execution. Unveiled at the 2026 Alibaba Cloud Summit on May 20, the model is built to manage intensive workflows such as iterative code modifications and extensive tool chaining that require sustained planning without human intervention.

Advanced Reasoning and Agentic Capabilities

Qwen3.7-Max utilizes an extended-thinking mode that generates a chain of thought before committing to a final answer. This internal reasoning process allows the model to plan, verify its work, and correct its course during execution. Because this approach generates significantly more output tokens than standard models—averaging 97 million tokens on Artificial Analysis benchmarks compared to the 24 million token average of other models—it is optimized for complex coding, debugging, and office workflow automation rather than simple, short-form tasks.
In internal testing conducted by Alibaba, the model demonstrated its agentic potential by autonomously performing over 1,000 tool calls and iterative code modifications to optimize a kernel. The company reported that this process improved inference speed by approximately 10x compared to the previous version. While these internal results are significant, they have not yet been independently verified.

Performance and Benchmarking

On the Artificial Analysis Intelligence Index, Qwen3.7-Max achieved a score of 56.6, securing the fifth position overall. This performance represents a 4.8-point improvement over its predecessor, the Qwen3.6 Max Preview. The gains are primarily concentrated in scientific reasoning, coding, and agentic capabilities, with notable increases in scores for benchmarks such as CritPt, Humanity’s Last Exam, and Terminal-Bench Hard.
However, the model’s performance on the AA-Omniscience benchmark indicates a shift in behavior. While its hallucination rate fell by 21.3 points, its raw accuracy dropped by 7.6 percentage points. This suggests the model is increasingly opting to state that it does not know an answer rather than attempting to recall facts, resulting in the lowest attempt rate among frontier models in that specific comparison. Developers relying on broad factual recall should account for this tendency toward abstention.

Technical Specifications and Availability

The model features a 1-million-token context window, a significant expansion from the 256K limit found in the Qwen3.6 Max Preview. This capacity allows the model to process large stacks of documents or mid-sized code repositories in a single request. Qwen3.7-Max is a text-only model, and Alibaba has also released a companion model, Qwen3.7-Plus-Preview, which supports vision and multimodal inputs.
Qwen3.7-Max is currently available as a preview build with closed weights. It is compatible with both OpenAI and Anthropic API specifications, allowing for integration into existing pipelines via Alibaba Cloud Model Studio. While pricing has not yet been announced, the model is positioned as a flagship tool for developers requiring advanced reasoning for long-horizon, multi-step autonomous tasks.