Qwen3.8-Max is The Next Chinese Open-Weights Assault on The AI Frontier
Chinese group Alibaba presented its new flagship AI model, Qwen3.8-Max, on Monday. It is available immediately through the QwenCloud platform. The model’s weights are set to be released next week. The accompanying test results show strengths in image, video and document processing – on coding, the model trails its US competitors.
Qwen3.8-Max has 2.4 trillion parameters, 95 billion of which are active per processing step. The model follows the “mixture of experts” principle: it consists of many sub-networks, only a fraction of which compute for any given request. The number of active parameters therefore determines how fast and how expensive it is to run.
Alibaba first announced the model on 19 July at the World AI Conference in Shanghai, at that point only as a preview. Alongside the official launch, a smaller variant called Qwen3.8-27B is due to become freely available next week. It would be the first time Alibaba has opened a model from its Max line. Until now the strongest models stayed closed, with only the mid-tier released.
What Alibaba is showcasing
Alibaba describes Qwen3.8-Max as a model for long, autonomously running tasks. The blog post cites several examples:
- In a single run lasting more than ten days, the model built a software project called “oh-my-cli” from scratch. After roughly 16 days it had produced 265 commits, 127 pull requests and 151 issues, according to the company. The repository is publicly viewable.
- In around 125 hours it reproduced a research paper on data selection for AI training, without any starter code, then developed improvements of its own.
- Designing a cryptographic chip automatically, it cut the number of required logic gates from 8,298 to 678 over roughly 500 working steps.
- In a simulation of a full trading year based on Taobao and Tmall data, it turned 100,000 yuan in starting capital into 416,252 yuan.
One technical detail stands out: Alibaba deliberately trained the model to run inside third-party agent environments as well. Besides its own QwenWork software, the company names Anthropic’s Claude Code, OpenAI’s Codex and the open-source projects OpenClaw and Hermes.
The test results
Alibaba has published a detailed comparison table. The reference models are Anthropic’s Claude Opus 4.8 and Claude Fable 5, OpenAI’s GPT-5.6 Sol, Google’s Gemini 3.1 Pro and Alibaba’s own previous generation.
For text and coding tasks, the table covers 31 tests. Qwen3.8-Max leads in six of them. In the coding category there are twelve tests, and the model leads in one: PaperBench. In most of the others, Claude Fable 5 or GPT-5.6 Sol comes out ahead.
| Test | Opus 4.8 | Fable 5 | GPT-5.6 Sol | Qwen3.7-Max | Qwen3.8-Max |
|---|---|---|---|---|---|
| Terminal Bench 2.1 | 84.6 | 84.6 | 88.8 | 74.5 | 86.6 |
| SWE-bench Pro | 69.2 | 80.0 | 64.6 | 60.6 | 67.7 |
| DeepSWE 1.1 | 59.0 | 70.0 | 73.0 | 21.6 | 56.6 |
| FrontierSWE | 70.0 | 88.8 | – | 40.7 | 73.5 |
| PaperBench | 80.3 | 88.8 | 90.5 | 64.8 | 93.0 |
| GPQA Diamond | 92.0 | 92.6 | 94.1 | 92.4 | 92.6 |
On two tests Alibaba developed itself, the new model also trails the Anthropic models. The same applies to RecreationBench, a test Alibaba introduces in the very same blog post: 51.7 points for Qwen3.8-Max against 56.1 for Claude Fable 5.
The model does lead on profession-related tasks: health questions (HealthBench 60.2), legal tasks (PLawBench 73.2), finance tasks (PRBench-Finance 58.3) and precise instruction following (IFBench 82.8 against 72.7 for GPT-5.6 Sol).
The gap to its own predecessor is substantial. On Terminal Bench 2.1 the score rose from 74.5 to 86.6 points, on DeepSWE from 21.6 to 56.6, and on FrontierSWE from 40.7 to 73.5.
Strengths in images, video and documents
The second table, covering the processing of images, video and documents, paints a different picture. It spans roughly 54 tests. Qwen3.8-Max leads in about two thirds of them, and in all seven document and office tasks.
The margin is widest in visual perception. On the Dense200 test, Qwen3.8-Max scores 87.0 points, Gemini 3.1 Pro 69.7 and Claude Fable 5 31.1. The model also leads on video understanding (VideoMME 90.4) and on physics and medical tasks involving images (HiPhO 90.0, SLAKE 90.8).
Caveats
Several points are worth bearing in mind when reading the numbers.
Claude Opus 5 is missing from the table. Anthropic released that model on 24 July. It currently tops the Intelligence Index compiled by the independent analysis firm Artificial Analysis with 61 points, ahead of Claude Fable 5 (60), GPT-5.6 Sol (59) and Kimi K3 (57). The comparison uses the older Opus 4.8 instead. The Chinese rival models Kimi K3 and GLM 5.2 do not appear in the table either.
Test conditions differ in places. For some tests Alibaba measured its own model itself and took published figures from other testing environments for the competition. On one test the company reports the better of two of its own results. For SWE-bench Pro, Alibaba says it corrected faulty tasks and tested all models on that revised version. OpenAI had found in a July audit that around 30 percent of the tasks in this test are flawed. The consequence, however, is that the figures can no longer be compared with the official leaderboard.
Seven of the tests are Alibaba’s own developments and cannot be verified externally. On three further tests, competing models grade the answers: PaperBench is judged by Claude Opus 4.6, two others by Gemini 3.1 Pro.
In a footnote, Alibaba also points out that the Claude Fable 5 results may include so-called fallbacks. Anthropic automatically routes requests on certain topics to a different model. That column therefore does not consistently measure the same model.
No independent measurements are available so far. Qwen3.8-Max is listed neither in the Artificial Analysis Intelligence Index nor on OpenRouter. There is no model card with figures on error rates or safety testing. Testers of the preview version also reported slow response times. It also remains unclear under which licence the weights are due to appear next week.
Pricing
Qwen3.8-Max costs 2 US dollars per one million input tokens and 6 dollars per one million output tokens. A token corresponds to roughly three quarters of a word. Cached input is billed at 0.25 dollars, and reading from an explicitly created cache at 0.17 dollars.
For comparison: Claude Opus 5 costs 5 and 25 dollars, Claude Fable 5 costs 10 and 50 dollars, GPT-5.6 Sol 5 and 30 dollars, Kimi K3 3 and 15 dollars, and GLM 5.2 1.40 and 4.40 dollars. Against its own predecessor, Qwen has become more expensive: Qwen3.7-Max cost 1.25 and 3.75 dollars. The 90 percent discount Alibaba offered during the preview phase ends with the regular launch.
The price per token, however, says little about the cost per task. Artificial Analysis measures that figure and arrives at a range from 0.05 dollars for DeepSeek V4 Pro through 0.86 dollars for Kimi K3 and 1.23 dollars for GPT-5.6 Sol up to 2.34 dollars for Claude Opus 5 and 3.15 dollars for Claude Fable 5. Models that “think” at length before answering generate many additional tokens, which are billed as output. On Qwen3.8-Max, that function is permanently active according to the documentation, with the highest setting as the default. No reliable cost-per-task figure exists for the model so far.
