Alibaba's Qwen team has released Qwen3.7-Max, an artificial intelligence model designed for extended autonomous operation that the company says achieved 35 hours of continuous execution on complex tasks. The model represents a departure from Alibaba's previous open-source releases, being offered only through paid APIs.

The model demonstrated its capabilities in an autonomous engineering task where it optimized an attention kernel on unfamiliar hardware over 35 hours. During this period, Qwen3.7-Max executed 1,158 tool calls and performed 432 kernel evaluations, achieving a 10.0x geometric mean speedup. The system operated without human intervention, diagnosing compilation failures and iteratively improving code.

Qwen3.7-Max features a 1-million-token context window and supports integration with existing agent frameworks, including compatibility with Anthropic's API protocol. On benchmark tests, the model scored 44.5 on Apex Math Reasoning, surpassing Claude Opus-4.6 Max's 34.5 and DeepSeek V4-Pro Max's 38.3. It also achieved high scores on other evaluations including Humanity's Last Exam and the MCP-Atlas coding benchmark.

The model is priced at $2.50 per million input tokens and $7.50 per million output tokens through Alibaba Cloud Model Studio. This positions it between budget-tier Chinese models and premium Western offerings like OpenAI's GPT-5.4 ($17.50 per million tokens) and Anthropic's Claude Opus 4.7 ($30.00 per million tokens).

The shift to proprietary licensing has drawn mixed reactions from the AI development community. While developers have praised the technical achievements, particularly the model's endurance capabilities, some have expressed disappointment over the loss of open-source access that characterized previous Qwen releases. The change aligns Alibaba with the commercial strategies of Western AI companies like OpenAI and Anthropic.

Separately, researchers have developed delta-mem, a memory enhancement technique that addresses AI agents' tendency to lose track of context during extended interactions. The system compresses historical information into a dynamically updated matrix that adds only 0.12% of parameters to existing models while outperforming alternatives that require 76.40% additional parameters. The technique allows models to maintain working memory without expanding context windows or relying heavily on retrieval systems.