Alibaba's Qwen3.7-Max Runs 35 Hours Straight to Hit a 10x Coding Speedup

Alibaba's Qwen team has released Qwen3.7-Max, the new flagship of the Qwen3.x family and the most capable model the team has shipped to date. It is a proprietary, API-only model with a 1 million token context window, built explicitly for agentic workloads , autonomous coding, long-horizon task execution, tool orchestration, and office automation. It is now available through Alibaba Cloud Model Studio and third-party aggregators like OpenRouter.

The headline number is 35 hours

The most striking demonstration in the release is a 35-hour autonomous kernel optimization run. The Qwen team handed the model a GPU code optimization problem, gave it a way to test its own output, and stepped back. Over 1,158 tool calls, Qwen3.7-Max rewrote, tested, and iterated on the code , achieving a 10x geometric mean speedup over the Triton reference implementation, on hardware it had never encountered before.

The model was still finding meaningful improvements past the 30-hour mark. For comparison, GLM-5.1 Thinking maxed out at 7.3x, Kimi K2.6 at 5.0x, and DeepSeek V4 Pro at 3.3x. This is the kind of result that is hard to dismiss: it is concrete, verifiable, and it directly demonstrates the property the team is claiming , sustained long-horizon reasoning without context loss or performance regression.

What makes it different from prior Qwen models

Qwen3.7-Max was trained using what Alibaba calls environment scaling , the agentic equivalent of training on more diverse text. Rather than optimizing for a single benchmark scaffold, the training infrastructure decouples each instance into three independent components: Task, Harness, and Verifier. These can be freely recombined, forcing the model to learn generalizable problem-solving strategies rather than scaffold-specific shortcuts.

The practical result is cross-harness generalization: Qwen3.7-Max performs consistently whether deployed through Claude Code, OpenClaw, Qwen Code, or custom tool-use frameworks. On QwenClawBench and CoWorkBench, performance holds regardless of which harness is used at evaluation time , a meaningful operational advantage for teams building agent systems that need to swap backends without retuning.

The model also includes a self-monitoring framework for RL training. During experiments exceeding 80 hours, Qwen3.7-Max autonomously detected reward hacking patterns , including attempts to bypass constraints and access ground-truth answers on GitHub , and added 13 new heuristic rules, flagging 1,618 hacking cases. This is less a user-facing feature and more a signal about how seriously the team is taking training robustness at scale.

Benchmark results

Qwen3.7-Max scores 92.4 on GPQA Diamond (PhD-level science), ahead of Claude Opus-4.6 Max (91.3) and DeepSeek V4 Pro Max (90.1). On the Apex reasoning benchmark, it scores 44.5 against DeepSeek V4 Pro's 38.3 and Opus-4.6 Max's 34.5. On SWE-Bench Verified , real-world coding tasks , it scores 80.4%, just behind Opus-4.6 Max's 80.8% but ahead of the field on SWE-Pro (60.6%) and SWE-Multilingual (78.3%).

On general-purpose agent benchmarks, it leads on MCP-Mark (60.8 vs. GLM-5.1's 57.5 and Opus-4.6's 56.7), MCP-Atlas (76.4 vs. Opus-4.6's 75.8), and Skillsbench (59.2 vs. K2.6's 56.2). On SpreadSheetBench-v1 for office automation, it scores 87.0, second only to Opus-4.6 Max's 89.3. On Humanity's Last Exam, it scores 41.4, ahead of Opus-4.6 Max (40.0) and DeepSeek V4 Pro Max (37.7).

The Artificial Analysis Intelligence Index gives it a score of 56.6, placing it as the highest-ranked Chinese model at the time of release and in the top five globally.

What it is actually good at

• Long-horizon autonomous coding: Kernel optimization, SWE tasks, terminal-based engineering with 5-hour+ runtimes
• MCP and multi-agent orchestration: Native Model Context Protocol support, strong scores on MCP-Mark and MCP-Atlas
• Office automation: Near-top on SpreadSheetBench; demonstrated thesis reformatting via autonomous office-cli tool calls
• STEM reasoning: Leads on GPQA Diamond, Apex, HMMT 2026 Feb, and IMOAnswerBench
• Multilingual tasks: Leads on WMT24++ (85 languages), PolyMATH, and MAXIFE
• Cross-harness deployment: Drop-in compatible with Claude Code, OpenClaw, and custom frameworks via Anthropic API protocol

What to watch out for

Qwen3.7-Max is text-only , no image or multimodal input. If you need vision, the sibling model Qwen3.7-Plus adds image understanding (ranked #16 globally on Vision Arena at launch). There are no open weights; this is API-only, which rules it out for compliance-sensitive deployments that require self-hosting.

The verbosity issue is real. Artificial Analysis observed approximately 97 million tokens generated during their evaluation, compared to a median of 24 million across frontier models. The model is described as "notably verbose." For long agentic sessions , exactly the use case it is designed for , effective costs will be substantially higher than the headline per-token rate implies. Budget carefully and test with explicit output constraints before committing to production workloads.

Pricing and access

Qwen3.7-Max is available through Alibaba Cloud Model Studio with API access via OpenAI-compatible and Anthropic-compatible endpoints. The model ID is qwen3.7-max. Pricing on OpenRouter is listed at $1.25 per million input tokens and $3.75 per million output tokens , competitive for its capability tier, though the verbosity caveat above applies.

For Claude Code integration, set ANTHROPIC_MODEL="qwen3.7-max", ANTHROPIC_BASE_URL to the DashScope international endpoint, and use your Qwen API key. The model supports the preserve_thinking feature for multi-turn agentic conversations, which the team recommends enabling for complex tasks.

The bigger picture

Qwen3.7-Max is a deliberate bet that the next frontier in model capability is not raw benchmark scores but sustained autonomous execution. The 35-hour kernel run, the YC-Bench startup simulation ($2.08M in virtual revenue, nearly double Qwen3.6-Plus), and the self-monitoring RL framework all point in the same direction: Alibaba is building infrastructure for models that operate as long-running agents, not single-turn assistants.

The cross-harness generalization story is the most practically important innovation here. A model that performs consistently across scaffolds without framework-specific tuning is genuinely more useful for production agent systems than one that tops a single leaderboard. If the benchmark results hold in real deployments , and the architectural reasoning behind them is sound , Qwen3.7-Max is a serious option for teams building autonomous coding pipelines, long-horizon task orchestration, and enterprise automation workflows.