Alibaba has introduced Qwen3.7-Max, a large language model positioned for long-running agentic work, code generation and scientific discovery. The company presents Qwen3.7-Max as its preferred text-only model; like other top-tier Qwen releases since late 2025, its model weights are not publicly released.
Capabilities and I/O limits
- Input: text up to 1 million tokens.
- Output: text up to 64,000 tokens.
- Throughput: 208.3 tokens per second (measured by Artificial Analysis at an unspecified reasoning setting).
- Features: reasoning, tool use, prompt caching, native compatibility with OpenAI’s and Anthropic’s API specifications, and the ability to retain reasoning text across turns.
Alibaba also released a multimodal companion, Qwen3.7-Plus-Preview, at the same time.
Benchmark performance
On the Artificial Analysis Intelligence Index — a composite of ten economically useful tasks — Qwen3.7-Max scored 56.6 on the reasoning measure, placing it roughly in the fifth-to-seventh range depending on how competing models’ reasoning levels are configured. It sits just behind Gemini 3.1 Pro Preview (57.2) and ahead of Google Gemini 3.5 Flash (55.3) in the reported comparisons.
Other Artificial Analysis findings include:
- AA-Omniscience (factual correctness measure that rewards correct outputs, penalizes incorrect ones, and does not count abstentions): Qwen3.7-Max, at an unspecified reasoning level, ranked sixth with a score of 14. The model posted a 23% hallucination rate — the lowest among frontier models tested — but that result was achieved in part by declining to answer more than half of prompts.
- Output speed: at an unspecified reasoning level the model reached 208 tokens per second, tying for third place with Gemini 3.5 Flash; GPT-OSS 120B (313 t/s) and GPT-OSS 20B (238 t/s) were faster.
Running the evaluations consumed roughly 97 million output tokens, substantially above the evaluators’ average of 35 million tokens.
Training approach and undisclosed details
Alibaba described Qwen3.7-Max’s reinforcement-learning approach at a high level. The company said it decoupled three components that are typically trained together in agent setups: the task to be performed, an agentic harness that calls tools, and a verifier that judges success. The model was trained on many combinations of task, harness and verifier to avoid learning behaviors specific to a single setup.
Alibaba did not disclose the model’s parameter count, internal architecture, training data or detailed training methods.
Internal test of agentic behavior and runtime optimization
Alibaba presented an internal, not yet independently validated test to demonstrate agentic capabilities. In that test the model purportedly optimized an attention kernel on hardware unseen during training: over 35 hours it made 1,158 tool calls and ran 432 kernel evaluations (test runs of candidate code). The generated code reportedly ran about ten times faster than a standard reference implementation.
Artificial Analysis has not yet run its own benchmark for long-running agentic tasks on Qwen3.7-Max.
Availability and pricing
- Free access via Qwen Chat (account required).
- API access through Alibaba Cloud Model Studio priced at $2.50 / $0.25 / $7.50 per million input / cached / output tokens respectively.
Strategic context
Qwen3.7-Max continues Alibaba’s move from open weights toward closed models. In addition to Qwen3.7-Max, Qwen3.6-Max-Preview and Qwen3.6-Plus have closed weights, while the weights for less capable Qwen3.6-27B and Qwen3.6-35B-A3B remain publicly available. Alibaba has also begun charging for Qwen Code, a command-line coding tool. These changes, following turnover in the Qwen team’s leadership, suggest Alibaba is prioritizing monetization of its top-tier models.
Why this matters
Measured by the Artificial Analysis Intelligence Index, Qwen3.7-Max ranks among the smartest Chinese LLMs and is one of the fastest models tested. However, the closed weights and lack of detailed technical disclosures limit independent evaluation and reproducibility.
Summary
Qwen3.7-Max represents a notable performance step for Alibaba’s LLM lineup: it delivers high throughput, agentic features and a low reported hallucination rate, while the company moves toward a business model that leverages closed, high-end models. Objective assessment of some claims will depend on future independent benchmarks and more detailed technical disclosure.
