The majority of AI models aren’t designed to execute in multiple steps and for a sustained period. For tasks such as running hundreds of code iterations or chaining calls to tools over several hours, a model with a training focus that is different from the current AI models will be required.
Qwen3.7 Max was officially announced by the Qwen team at 2026 Alibaba Cloud Summit, on 20 May. Two preview versions of Qwen3.7 appeared quietly on Arena AI leaderboard without any press release or official API announcement.
The release of two preview models at the same time
Alibaba previewed two models simultaneously: Qwen3.7-Max-Preview and Qwen3.7-Plus-Preview. According to LM Arena, they ranked respectively 13th and 16th worldwide in terms of text capability.
Alibaba is the 6th lab in Text Arena. Qwen3.7 Max-Preview was ranked 13 overall. Qwen3.7 Plus-Preview placed #16 in Vision Arena. Alibaba is the #5 laboratory. Model rank and lab rank are two separate numbers.
The Qwen3.7 Plus-Preview preview is a balanced high-performance version that focuses on logic and reasoning. Its toolchain will be opened gradually in the future. This version supports multimodal input and visuals. Qwen3.7 is the flagship text-only reasoning. The Qwen3.7 Max model is covered in this article, since it’s the one Alibaba officially announced.
Why is Qwen3.7 Max Designed for?
Alibaba Qwen’s team describes Qwen3.7 Max as their most sophisticated and comprehensive agent to date. Model is proprietary, closed-weight. It can handle coding and bugging, automated office workflows, and large-scale tasks with hundreds, if not thousands, of steps.
Long-Term Thinking Mode
The Qwen3.7 Max is a logic model. The model generates a chain of thought first — an internal sequence of steps where it plans, checks its work, and corrects course before committing to a final answer. On interfaces like Qwen Chat, this shows up as a ‘Thinking’ mode you can switch on to see the model’s reasoning trace.
The output tokens of reasoning models are significantly higher than those produced by standard completions. Artificial Analysis evaluated the Intelligence Index using Qwen3.7. Max. It generated around 97 million tokens, compared to an average of 24 million For models based on this benchmark. This overhead increases latency for short tasks or simpler ones without increasing output quality. The extended-thinking model is most useful for multi-step plans, code refactoring or large agent chains.
Contextual Window
Model features 1M context windows, up to 256K in Qwen3.6. The model supports only text input and outputting. The pricing has yet to be announced. Alibaba Cloud charges $1.30/$7.80 for each million tokens of input/output.
One context window that can store a thousand tokens is enough to hold an entire mid-sized repository of code or many documents. As the context window grows, models tend to become less reliable. Qwen3.7 Max is still not available for independent long context testing.
Benchmark Results
The Artificial Analysis Intelligence Index ranked Qwen3.7 Max at 56.6, ranking it as the fifth best overall. It is up 4.8 points from its predecessor Qwen3.6 Max Preview (51,8) and ahead of Google Gemini 3.5 Flash’s (55.3). GPT-5 (55.2), Claude Opus 4.7 (57.3), and Gemini 3.1 Preview (57.2), still top the list.
The Intelligence Index v4.0 contains ten assessments, such as GDPval-AA (Average), Terminal-Bench-Hard, SciCode, AA Omniscience, Humanity’s Last Exam and GPQA.
The improvements over Qwen3.6 are not consistent. Index improvements are concentrated primarily in three areas: scientific reasoning and agentic capabilities. CritPt jumped from 3.7% to 134%, Humanity’s last exam jumped by 9.2 percentage points and Terminal Bench Hard rose 6.9 percentage points. GDPval-AA gained 42 Elo (from 1504 to 1546). Other benchmark scores are mostly flat in comparison to Qwen3.6 Maximum Preview.
The Index is a complex index. In AA Omniscience, the raw accuracy for Qwen3.7 Max actually fell 7.6 percentages points (from 36.7% to 30%.1) while its hallucination ratio dropped 21.3 percentage points (from 42.2% to 22.9%). Model is choosing what to say “I don’t know” More facts are not necessarily better. Its rate of attempts fell from 67.3 to 48.0%. That was the lowest attempt rate among all frontier models in this comparison. The AA Omniscience Benchmark rewards correct answers but penalizes hallucinations. There is no penalty if you refuse to answer. If your use case relies on broad facts being remembered, then this limitation is important to consider when testing against workload.
Qwen3.7 Max-Preview was ranked 13th overall in Text Arena with an Elo of 1475. The top ten categories are: #7 for Math, #9 for Expert Prompts and IT and #10 for Coding.
All benchmarks are tentative. The model carries a ‘Preview’ mode, indicating Alibaba considers it an early build.
Agentic Performance — Internal Test
Alibaba’s internal test of a chip platform saw the model perform more than 1,000 calls to tools and make iterative changes in code. This was done for optimizing a kernel. Alibaba said the inference process increased speed roughly by 10x when compared to the old version.
Marktechpost’s Visual Explainer
The Key Takeaways
- Alibaba has released two Qwen3.7 models as previews: Plus (multimodal) and Max (text/reasoning).
- Qwen3.7-Max scored 56.6 on the Artificial Analysis Intelligence Index, ranking #5 overall — a 4.8-point gain over Qwen3.6 Max Preview.
- Text only; no images.
- On AA-Omniscience, raw accuracy dropped while abstention rose — worth testing for knowledge-recall use cases.
- Alibaba has only conducted internal tests on the model, and no external verification.
Check out the Technical details. You can also find out more about the following: Docs. Also, feel free to follow us on Twitter Join our Facebook group! 150k+ ML SubReddit Subscribe now our Newsletter. Wait! Are you using Telegram? now you can join us on telegram as well.
You can partner with us to promote your GitHub Repository OR Hugging Page OR New Product Launch OR Webinar, etc.? Connect with us

