Qwen3.7-Max review: Alibaba's agent flagship that ran for 35 hours

*Qwen3.7-Max (qwen3.7-max) is built for long jobs you hand to an agent and walk away from. It is not chasing the top of the general-chat leaderboards; it is tuned to keep working, call tools and recover from its own errors over hours.

What it is: Qwen's closed flagship

Alibaba announced Qwen3.7-Max in May 2026 as the top of the Qwen3.7 series. A common misreading: this model is not open weights. Qwen publishes many open models, but the Max tier is closed.

• Context: about 1M tokens as published by Alibaba, up to 64K tokens of output
• Input and output: text only
• Positioning: a flagship optimized for agent work

On FastMetal, Qwen3.7-Max runs on GMI Cloud, which serves 262,144 tokens of context, about a quarter of Alibaba's figure. That is still a large codebase, but not the whole-repository-in-one-prompt scale the 1M number suggests.

Benchmarks: ahead of Claude Opus 4.7 on agent work

On reported figures it beats Claude Opus 4.7 on agentic coding benchmarks:

It was also reported at 80.4 on SWE-bench Verified and 78.3 on SWE-Multilingual. On Artificial Analysis's Intelligence Index v4.0 it scored 56.6, inside the global top five and reported as the highest-ranked Chinese model.

The 35-hour run

Alibaba's technical write-up describes Qwen3.7-Max running a CUDA kernel optimization task autonomously for 35 hours. Over that run it made 1,158 tool calls and 432 kernel evaluations, diagnosed its own compile errors, kept iterating, and ended with a 10x geometric-mean speedup.

What that shows is stamina more than raw intelligence: holding onto state across a very long session. Benchmarks that score a single exchange do not capture it.

Cost

With long-running agents, total tokens matter more than the per-token rate. Thirty-five hours of work consumes thirty-five hours of tokens, so estimate per task, not per million.

On FastMetal, Qwen3.7-Max is billed per million input and output tokens; current rates are on the model page. It reasons before answering, and those reasoning tokens bill as output. The balance is prepaid and requests stop when it runs out, so a runaway agent cannot run up an open-ended bill.

GMI Cloud has not published a training or retention policy for inference traffic, so FastMetal lists this route's data policy as unknown. Keep that in mind before sending it anything sensitive.

Non-English use

Qwen3.7-Max is multilingual and handles general tasks in other languages. Note that SWE-Multilingual's 78.3 measures coding across programming languages and says nothing direct about writing quality in a natural language. We found no language-specific benchmarks, so test on your own prompts if register or terminology matters.

Good fit, poor fit

*### Good fit

• Long, multi-step autonomous agent runs
• Automation with heavy terminal use and many tool calls
• Batch development tasks you want to run without someone watching

*### Poor fit

• Images or audio (it is text only)
• Requirements that call for open weights (it is closed)
• Latency-first workloads that return many short replies

FAQ

*### Q. Is Qwen3.7-Max open source?

No. The Qwen family includes many open-weight models, but the Max tier is closed.

*### Q. What sets it apart from other models?

It is optimized for long autonomous runs. It was reported ahead of Claude Opus 4.7 on Terminal-Bench 2.0, SWE-Bench Pro and MCP-Atlas.

*### Q. What context window does FastMetal serve for Qwen3.7-Max?

262,144 tokens, from GMI Cloud. Alibaba's published figure is about 1M.

Try it on FastMetal

Give it room in max_tokens, since its reasoning counts against the output budget.

curl https://api.fastmetal.ai/v1/chat/completions \
  -H "Authorization: Bearer $FASTMETAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen3.7-max",
    "max_tokens": 16000,
    "messages": [{"role": "user", "content": "Find the cause of this failing test and propose a fix: ..."}]
  }'

The same API key and balance work for Claude and the rest of the catalog, so you can change model and compare. The full list is in the model catalog.