Alibaba Qwen3.7-Max: 35 hours of autonomous code optimization on a custom chip
Date: 2026-05-23 Source: The Decoder (Jonathan Kemper), via Alibaba Qwen team. Links: The Decoder Raw source: raw/rss/2026-05-23-the-decoder-alibabas-latest-ai-model-ran-autonomously-for-35-hours.md
TL;DR
Alibaba released Qwen3.7-Max, a proprietary model built specifically for long-horizon autonomous agent tasks. The headline demo: 35 hours of continuous autonomous operation optimizing code for Alibaba's own custom AI chip. On standard benchmarks it matches Claude Opus 4.6 and beats Chinese rivals DeepSeek V4 Pro and Kimi K2.6. The team also demonstrated it steering a four-legged robot. The 35-hour demo is the differentiator: it is one of the longest verified autonomous-execution windows for a coding agent reported by a major lab.
Why this matters
Three threads cross at this release.
Long-horizon agentic capability is the new benchmark. The SR^2AM paper today (the paper that adds learned self-regulation to decide when and how deeply to plan, cutting reasoning tokens by 25-95% while remaining competitive with 685B-1T parameter systems) is the academic-side parallel. Qwen3.7-Max running for 35 hours autonomously is the production-side test of the same thesis: can a learned planner schedule itself over horizons that exceed any single context window? The 35-hour number is meaningful because it sits well beyond the Anthropic Opus 4.5 vs 4.6 single-task "all-day coding" demos, and approaches the territory where the model is essentially running its own daily standup.
The China-stack is consolidating around custom silicon. Qwen3.7-Max optimizing code "for its own custom chip" is the load-bearing detail. Alibaba has been building T-Head chips since 2019 (Hanguang for inference, Yitian for general-purpose). A model that can autonomously tune the kernel code that runs on its sibling chip closes a loop NVIDIA does not yet have. The wiki's SemiAnalysis Cerebras IPO piece (2026-05-13) tracked how custom-silicon-plus-tuned-model is the structural pattern: this is the same pattern visible from Hangzhou.
Open vs proprietary tension inside China. Qwen3.7-Max is explicitly proprietary, breaking the Qwen team's recent open-weight tradition. The same week, DeepSeek made its V4-Pro discount permanent (signaling commitment to open weights via accessible pricing) and Kimi K2.5 went into production via Cursor / Fireworks (commercial open-weights). Alibaba going proprietary on the long-horizon-agent SKU while keeping smaller Qwen variants open is a clear product-line segmentation move: open weights for the developer mindshare, proprietary for the high-margin agent product.
Connecting to the wiki
The wiki has been tracking the long-horizon coding-agent thread since the 04-22 Claude Code architecture page, the 05-22 Anthropic profitability story (coding-agent revenue driving the first-profitable-quarter projection), and the 05-23 Composer 2.5 release (Kimi K2.5 + Cursor stack hitting #3 on the Artificial Analysis coding-agent index at 1/10 the cost).
This is now a four-way race in coding agents: Anthropic Claude Code / Mythos, OpenAI Codex 5.5, Cursor Composer 2.5 (Kimi-backed), and Alibaba Qwen3.7-Max. The Chinese stack just demonstrated the longest continuous autonomy window of any of them.
Industrial implication
If the 35-hour run reproduces across third-party tests (independent of Alibaba's hand-picked task), the practical implication is that "all-day autonomous coding" stops being an aspirational marketing line and becomes a feature checkbox. The downstream consequence: enterprise procurement of coding agents shifts from "is Claude or GPT-5 better at the prompt?" to "which model can I leave running unsupervised for the longest?" This favors whoever has the strongest replay-buffer training data for multi-turn agent trajectories.
For competition: NVIDIA's strategic moat is that no one else's silicon runs the same workload as fast. Alibaba is testing whether a model-plus-chip co-design loop (Qwen3.7-Max tuning T-Head kernels) can erode that moat from the China side. NVIDIA's next move is presumably to make CUDA kernel optimization the kind of task its own published models can do end-to-end on Rubin or Vera silicon.
Gaps
- The 35-hour task is one demo. We do not know the failure mode distribution, the cost per hour, or whether the model maintained coherence across the full 35 hours or hit recovery loops.
- "Matches Claude Opus 4.6" is benchmark-language. The specific benchmarks are not yet reported in detail.
- Whether Qwen3.7-Max will be accessible outside Alibaba Cloud, and at what price tier, is undisclosed.
Worth watching
- Whether any non-Alibaba team reproduces a 24+ hour autonomous agentic-coding run within 60 days.
- Whether NVIDIA responds with a kernel-tuning agent demo at GTC Taipei (June 1) or at SIGGRAPH.
- Whether Qwen3.7-Max coding benchmark numbers land within 5% of Claude Opus 4.6 on independent third-party evaluations (90 days).
- Whether Alibaba ships a smaller open-weight Qwen3.7 variant for developer ecosystem (60 days).