Released May 20, 2026 at the Alibaba Cloud Summit in Hangzhou, Qwen3.7 Max scored 56.6 on the Artificial Analysis Intelligence Index — placing it #5 globally at launch and making it the highest-ranked Chinese AI model ever recorded. It leads every competitor on competition mathematics, tops HLE, and ran a 35-hour autonomous coding session. Here is what you need to know.
What Changed From Qwen3.6
Qwen3.7 Max is a closed-weight, API-only model available through Alibaba Cloud DashScope and OpenRouter. It is the text-only flagship, joined by Qwen3.7-Plus — a cheaper multimodal variant with vision input.
The jump from Qwen3.6-Plus is significant. On Terminal Bench 2.0-Terminus, Qwen3.7 Max scores 69.7 vs Qwen3.6-Plus's 61.6. On the YC-Bench startup simulation, it doubled revenue generation (2.08M vs 1.05M). This is not an incremental update — it is a different tier of model.
Benchmark Breakdown
| Benchmark | Qwen3.7 Max | Notable Comparisons |
|---|---|---|
| AA Intelligence Index | 56.6 | Between Gemini 3.5 Flash and Claude Opus 4.7 |
| HMMT 2026 (math) | 97.1% | Beats Claude Opus 4.6 (96.2%), DeepSeek V4 (95.2%) |
| HLE (Humanity's Last Exam) | 41.4% | Beats Opus 4.6 (40.0%), DeepSeek V4 (37.7%) |
| GPQA Diamond | 92.3% | Graduate-level scientific reasoning |
| IMOAnswerBench | 90.0% | DeepSeek V4 at 89.8% |
| PolyMATH | 86.5% | Opus 4.6 at 80.2% |
| WMT24++ (multilingual) | 85.8% | Opus 4.6 at 82.7% |
| Apex benchmark | 44.5 | Opus 4.6 at 34.5, DeepSeek V4 at 38.3 |
The math and reasoning story is remarkable. Qwen3.7 Max leads every competitor on HMMT 2026 competition mathematics and scores above 90% on GPQA Diamond — graduate-level science questions. Its multilingual translation (WMT24++) also tops the field.
The 35-Hour Autonomous Coding Run
The standout demo from the launch was a 35-hour autonomous coding session. The model fired 1,158 tool calls during a GPU kernel optimization task, achieving a 10x speedup over the standard Triton reference implementation. This is the kind of sustained, multi-step engineering work that requires holding context across hundreds of iterations — exactly the agentic use case Alibaba is targeting.
Pricing and Access
- Input: $2.50 per million tokens
- Output: $7.50 per million tokens
- Cached input: $0.25 per million tokens (90% discount)
- Context window: 1 million tokens
- Max output: 65,536 tokens per request
That pricing places it above DeepSeek V4 Pro ($0.435/$0.87) but below Claude Opus 4.7 and GPT-5.5. The 1M context window matches Gemini 3.1 Ultra and exceeds most competitors.
Availability: Alibaba Cloud DashScope, Alibaba Cloud Model Studio, and OpenRouter. No open weights — API only.
Where It Trails
The model is notably verbose. Independent testing reports it generates roughly 97 million tokens per evaluation run compared to a median of 24 million — meaning cost at scale can be significantly higher than headline pricing suggests.
It also sits below the very top tier on the Intelligence Index. GPT-5.5 and Claude Mythos both scored higher on overall composite benchmarks. And being closed-weight limits fine-tuning and local deployment options that the Qwen open-source community has come to expect.
The Bigger Picture
Qwen3.7 Max represents China's strongest entry yet in the global AI race. It proves that Chinese labs can compete at the frontier on reasoning and agentic tasks — not just efficiency and cost. The question now is whether Alibaba opens the weights, as they have with most previous Qwen releases, or keeps this one closed to differentiate their cloud platform.
For developers, the combination of 1M context, strong agentic performance, and competitive pricing makes Qwen3.7 Max worth testing — especially for long-horizon coding, multilingual, and math-heavy workloads.
Status: Preview. Available for testing on Qwen Chat with thinking mode enabled. No official model card or public benchmark release yet — treat all numbers as preliminary until Alibaba publishes final results.