๐ผPandailyโขStalecollected in 1m
Alibaba Qwen 3.7 Max Completes 35-Hour Autonomous Task
๐กQwen 3.7 Max proves its reliability for long-running AI agents with 1,158 successful tool calls.
โก 30-Second TL;DR
What Changed
Sustained 35-hour autonomous operation capability
Why It Matters
This performance benchmark suggests that Qwen 3.7 Max is highly suitable for complex agentic workflows that require long-term planning and tool utilization.
What To Do Next
Evaluate Qwen 3.7 Max for your next agentic project that requires long-duration task execution and frequent tool interaction.
Who should care:Developers & AI Engineers
Key Points
- โขSustained 35-hour autonomous operation capability
- โขSuccessfully processed 1,158 tool calls during the run
- โขDemonstrates high reliability for complex, long-running AI agents
๐ง Deep Insight
Web-grounded analysis with 13 cited sources.
๐ Enhanced Key Takeaways
- โขThe 35-hour autonomous task involved optimizing a hardware-based attention kernel for the open-source inference software SGLang on Alibaba's custom T-Head-ZW-M890 accelerators, a chip architecture the model had not encountered during training.
- โขDuring the task, Qwen 3.7 Max achieved a 10x geometric mean speedup over the reference implementation, significantly outperforming competitor models like GLM 5.1 (7.3x speedup) and Kimi K2.6 (5x speedup) in the same optimization setup.
- โขQwen 3.7 Max is a proprietary model, available exclusively through the Alibaba Cloud Model Studio API, marking a strategic shift from Alibaba's previous approach of releasing flagship Qwen models as open source.
- โขThe model is specifically designed for agent-based tasks, targeting use cases such as coding agent work (from front-end prototypes to multi-file projects), automating office tasks with external tools, and running autonomously for long durations across various agent frameworks.
- โขIt features a 1-million-token context window, a substantial increase from its predecessor Qwen3.6 Max Preview's 256K limit, and supports OpenAI- and Anthropic-compatible interfaces for broader integration.
๐ Competitor Analysisโธ Show
| Feature/Model | Alibaba Qwen 3.7 Max | Claude Opus 4.6 Max / 4.7 | DeepSeek V4 Pro Max | Kimi K2.6 Thinking | GLM-5.1 Thinking |
|---|---|---|---|---|---|
| Availability | Proprietary, API-only via Alibaba Cloud Model Studio | API (Anthropic) | API (DeepSeek) | API (Moonshot AI) | API (z.ai) |
| Context Window | 1 Million tokens | Up to 200K tokens (Claude generally) | Not specified, competitive | Not specified, competitive | Not specified, competitive |
| Pricing (per 1M tokens) | Input: $2.50, Output: $7.50 | Not specified, but noted as potentially costing less in practice despite higher rates | Not specified | Not specified | Not specified |
| SWE-Verified Benchmark | 80.4 | 80.8 | 80.6 | Not specified | Not specified |
| Apex Math Reasoning | 44.5 | 34.5 (Opus-4.6 Max) | 38.3 (DeepSeek V4-Pro Max) | Not specified | Not specified |
| Kernel Optimization Speedup (35hr task) | 10.0x geometric mean | Not specified, but noted as often superior for correctness-critical engineering | 3.3x | 5.0x | 7.3x |
| Key Focus | Agentic coding, long-horizon tasks, office automation, cross-framework consistency | Coding, analysis, long-document understanding, AI safety | High-performance open-source models, coding, reasoning | Large language models | Large language models |
๐ ๏ธ Technical Deep Dive
- Qwen 3.7 Max is a proprietary reasoning model, available exclusively via API, with text-only input and output capabilities.
- It features a substantial 1-million-token context window, designed to support long-horizon reasoning and prevent performance degradation over extended tasks.
- The model was demonstrated optimizing an attention kernel on Alibaba's T-Head ZW-M890 PPU, a custom AI chip platform, without prior training exposure to its architecture.
- Qwen 3.7 Max supports native compatibility with mainstream agent harnesses, including OpenAI- and Anthropic-compatible interfaces like Claude Code, OpenClaw, and Qwen Code.
- It incorporates self-monitoring mechanisms to detect undesirable behavior and 'reward hacking' during its own training process, writing new detection rules and flagging cases.
- While specific architectural details for 3.7 Max are not fully disclosed, previous Qwen models like Qwen 3 and Qwen 3.5 have utilized Mixture-of-Experts (MoE) architectures and were trained on trillions of tokens across numerous languages.
๐ฎ Future ImplicationsAI analysis grounded in cited sources
Alibaba's pivot to proprietary, API-only models for its most advanced Qwen versions will intensify competition with Western AI giants in the enterprise agent market.
By offering Qwen 3.7 Max exclusively via API and focusing on agentic capabilities, Alibaba is directly aligning with the commercial strategies of leading Western AI labs, moving beyond its previous open-source emphasis for flagship models.
The demonstrated long-duration autonomous task capability will accelerate the adoption of AI agents for complex, multi-step enterprise workflows.
Qwen 3.7 Max's ability to maintain coherence and continuously optimize over 35 hours for a real-world engineering task showcases a new level of reliability for AI agents, making them more viable for critical business operations.
The integration of AI models with custom hardware, as seen with Qwen 3.7 Max and the T-Head ZW-M890 chip, will become a key differentiator in the AI industry.
Alibaba's 'full-stack pitch' of model and silicon together suggests a trend towards vertical integration to optimize performance and efficiency for specialized AI tasks, potentially creating a competitive advantage.
โณ Timeline
2023-04
Alibaba launched a beta of Qwen (Tongyi Qianwen).
2023-09
Qwen opened for public use after regulatory clearance.
2024-06
Qwen2 was released.
2024-09
Qwen2.5 arrived.
2025-04
Qwen3 launched, introducing hybrid thinking mode and supporting 119 languages.
2026-05-20
Qwen3.7-Max unveiled at Alibaba Cloud Summit.
๐ Sources (13)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
๐ฐ
Weekly AI Recap
Read this week's curated digest of top AI events โ
๐Related Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: Pandaily โ