--- title: "Alibaba's Qwen3.7-Max Becomes the Highest-Ranked Chinese AI Model" subtitle: "Alibaba's Qwen3.7-Max scores 56.6 on the Intelligence Index, becoming the top-ranked Chinese model — but its hallucination gains come with a catch" url: "https://alphasignal.ai/news/alibaba-s-qwen3-7-max-becomes-the-highest-ranked-chinese-ai-model" author: "AlphaSignal Newsroom" author_url: "https://alphasignal.ai/editorial-team" company: "Artificial Analysis" company_url: "https://alphasignal.ai/company/artificialanlys" published: "2026-05-21T08:14:28.000Z" updated: "2026-06-16T21:18:37.821Z" type: "News" topics: - "Llms" - "Benchmarks" subtopics: - "Long Context" - "Hallucinations" upvotes: 1204 read_time: "5 min" source_url: "https://artificialanalysis.ai/models/qwen3-7-max" comments: 0 --- # Alibaba's Qwen3.7-Max Becomes the Highest-Ranked Chinese AI Model > Alibaba's Qwen3.7-Max scores 56.6 on the Intelligence Index, becoming the top-ranked Chinese model — but its hallucination gains come with a catch Alibaba has released [Qwen3.7-Max](https://artificialanalysis.ai/models/qwen3-7-max), its new proprietary flagship reasoning model, and the numbers are worth paying attention to. The model scores 56.6 on the Artificial Analysis Intelligence Index, a 4.8-point gain over its predecessor Qwen3.6 Max Preview (51.8), placing it fifth overall and making it the highest-ranked Chinese model on the leaderboard. GPT-5.5 (60.2), Claude Opus 4.7 (57.3), and Gemini 3.1 Pro Preview (57.2) still lead the pack , but Alibaba has never been this close. ## From open to closed, again Since Qwen2.5 Max in January 2025, Alibaba has followed a consistent strategy: release Max and Plus variants as closed-weight proprietary models while keeping the rest of the Qwen line open. Qwen3.7-Max continues that pattern. Qwen3.7-Max is not open source and not open-weight , you cannot download it and you cannot run it on your own GPUs. The leading open-weights Qwen on the Intelligence Index remains Qwen3.6 27B (Reasoning, 45.8), while the leading open-weights MoE Qwen is Qwen3.5 397B A17B (Reasoning, 45.0). The shift is deliberate. Alibaba updated its flagship model for long-running agentic work, pushing it into the top rank among LLMs built in China. Alongside Qwen3.7-Max, Alibaba also released Qwen3.7-Plus, the multimodal sibling with vision support. ## Where it actually improved The 4.8-point jump on the Intelligence Index is real, but it is not evenly distributed. The gains are concentrated in three areas: - **Scientific reasoning:** Most of the Index gains are concentrated in scientific reasoning, agentic capability, and coding. CritPt (physics reasoning) jumped 9.7 percentage points, from 3.7% to 13.4%. - **Hard science benchmarks:** Qwen3.7-Max posts 92.4 on GPQA Diamond and 97.1 on HMMT 2026 February, the highest score in its comparison group. - **Agentic coding:** Terminal-Bench Hard improved 6.9 percentage points to 50.8%, and it lands 80.4 on SWE-Verified for software engineering tasks. Qwen3.7-Max is genuinely competitive at the frontier on coding-agent benchmarks , its SWE-Pro and Terminal-Bench numbers are within noise of Claude Opus 4.7 and GPT-5.5, and ahead of DeepSeek V4 Pro on the same suite. ## The hallucination number needs a footnote The most headline-grabbing stat is Qwen3.7-Max's hallucination rate. Qwen3.7-Max reports the lowest hallucination rate among frontier models at 22.9%. That sounds like a clean win , but the mechanism matters. On AA-Omniscience, Qwen3.7-Max's raw accuracy actually dropped 7.6 percentage points (from 37.7% to 30.1%), while its hallucination rate fell 21.3 points (from 44.2% to 22.9%). The model is not recalling more facts. It is refusing to answer more questions. The model's attempt rate fell from 67.3% to 48.0%. It refuses to answer more often, which lowers wrong answers but also lowers usefulness on edge cases , a trade-off that matters if you're plugging it into an agent that needs to push through ambiguity. ## How it was trained Alibaba gave a high-level description of the training approach. Alibaba reports that the model can sustain over 35 hours of continuous execution, working on a single prompt task for more than a full day. The reinforcement learning setup is notable: in an internal test on a new chip platform, the model autonomously performed more than 1,000 tool calls and iterative code modifications to optimize a key kernel, with Alibaba claiming the process improved inference speed by roughly 10x compared with the previous version. The training design specifically targets generalization across agent setups. Alibaba separated three components that are typically coupled in agent training: the task to be performed, the agentic harness that calls tools, and the verifier that decides whether the system succeeded. Training across many combinations of these three was meant to prevent the model from learning tricks specific to a single evaluation setup. That said, the agentic claim is based on an internal test not yet validated by independent benchmarks. ## Speed, context, and what it costs On the infrastructure side, Qwen3.7-Max has some genuinely strong numbers. Alibaba increased the context window from 256K tokens on Qwen3.6 Max to 1M tokens on Qwen3.7-Max, bringing it in line with Claude Opus 4.8 and other contemporary flagship models. Output speed clocks in at roughly 206 tokens per second, placing it third overall among frontier models. Pricing is now confirmed: API pricing is $2.50 per million input tokens and $7.50 per million output tokens, with cached input dropping to $0.25 per million. That makes Qwen3.7-Max roughly half the price of Claude Opus 4.8 on input and less than a third on output. There is a catch on cost though: the model used 97M output tokens to run the Intelligence Index evaluation , nearly three times the 35M average across frontier models , so real-world costs will run higher than the rate card suggests for complex tasks. ## Where to use it, and where not to Alibaba positions Qwen3.7-Max as its preferred model for text-only work like coding and scientific discovery. Access is available through: - Alibaba Cloud Model Studio (API live as of May 19), with cross-listings on OpenRouter, Together AI, and Qubrid AI. - Free via Qwen Chat with an account. The model is text-only , no image input. For multimodal tasks, use Qwen3.7-Plus-Preview instead, which supports vision input. The high abstention rate is the main thing to test before committing. In real use, you will get answers like "I don't know" more often than with other models, and it will refuse to answer some edge cases where other AI will at least try. For agentic pipelines that need to push through uncertainty, that behavior needs to be validated on your specific workload before deployment. ## The bigger picture Qwen3.7-Max is the clearest signal yet that Alibaba is competing for revenue, not just research prestige. The move to closed weights at the flagship tier, combined with the start of paid access to Qwen Code, points to a deliberate monetization strategy. Chinese labs are competing among themselves on who can be closest to the US-led frontier , and with Qwen3.7-Max, Alibaba has pulled ahead of every other Chinese lab on the most comprehensive third-party benchmark available. The gap to OpenAI, Anthropic, and Google is narrowing, even if it has not closed. ## Sources - [X post by @artificialanlys](https://x.com/artificialanlys/status/2057374452883788196) - [Qwen3.7 Max - Intelligence, Performance & Price Analysis](https://artificialanalysis.ai/models/qwen3-7-max) - [Alibaba's Latest Proprietary Model, Qwen3.7-Max, Challenges U.S. Rivals](https://www.deeplearning.ai/the-batch/qwen3-7-max-adds-speed-and-power) - [Alibaba introduces Qwen3.7-Max as next-gen AI agent model](https://technode.com/2026/05/21/alibaba-introduces-qwen3-7-max-as-next-gen-ai-agent-model/) - [Qwen3.7-Max Review 2026: Benchmarks, Pricing, Verdict](https://felloai.com/qwen-3-7-max-review/) - [Alibaba's Qwen 3.7 Max Becomes Highest-Placed Chinese Model On Artificial Analysis Index](https://officechai.com/ai/qwen-3-7-max-benchmarks/)