Qwen shipped Qwen3.8-Max today, and the company chose to lead its own announcement with a coding agent that ran unsupervised for roughly 16 days, filing 265 commits, opening 127 pull requests, and closing 151 issues without a human approving any of it. The release, dated 3 August 2026, is general availability. AI Insiders covered this model on 21 July as a preview build with limited access; today it is live for anyone with a QwenCloud account, and Qwen says the open weights, its first release of a Max-class model under an open license, arrive next week rather than today. Every other claim in the announcement, benchmark table included, is Qwen’s own account of what its 2.4-trillion-parameter model did, and none of it carries outside verification yet.

The most measurable of the three demonstrations targeted chip design. Qwen set the model loose on a cryptographic hardware accelerator with no reference design and no human intervention, judged only by whether it held bit-exact correctness at four separate bit widths (4, 6, 8, and 16) while it shrank the synthesized gate count. Over roughly 500 turns and 71 evaluations, the first working version at 8,298 gates fell to 678, with a single algorithmic rewrite (swapping a modulo divider for an iterative shift-subtract approach) doing more than 80 percent of that reduction on its own. Run through place-and-route, the design’s die footprint dropped from 106 by 106 micrometres to 46 by 46, an 81 percent cut in physical area, while hitting timing closure at 500 MHz.

The second run asked the model to reproduce a research paper and then beat it. Qwen assigned it a paper called “Unified Data Selection,” on how to pick the most valuable training examples when there is far more data available than compute to train on, the kind of problem researchers hit constantly. Starting from nothing but the paper text and a GPU allocation, the model worked for about 125 hours, produced close to 7,600 lines of working code, and ran 33 rounds of training. It spent the opening 37 hours reconstructing the original pipeline and confirming the paper’s six central findings, matching its claim that smart selection beats random sampling by 7.7 percentage points on the AIME24 math benchmark. It then spent the remaining 88 hours proposing and testing, 18 self-generated improvement ideas spread over four rounds, and landed on a method that scored 2.71 points higher than the paper’s own approach, moving from a 49.58 percent baseline to 52.29 percent.

The third case, the 16-day run named in the lede, built a self-evolving coding harness called oh-my-cli from a blank repository. User bug reports, established coding conventions, and the harness’s own test output all feed a single engineering loop: issues move through queued, claimed, and active states, and a pull request merges only once end-to-end tests and CI checks pass. Qwen has posted the trace publicly on GitHub, which at least gives outside engineers something to inspect rather than take on faith.

None of that changes the fact that every one of these three results comes from Qwen describing its own model’s performance, with no outside lab, competition, or auditor confirming the claims independently. The 21-benchmark comparison table in the same release is a set of tests Qwen chose to publish, not a neutral panel’s selection, and it should be read that way. It is also worth stating plainly that Qwen’s claim to be first-to-open-source a Max-class model is not true yet: the weights are promised for next week, and a promised release is not a shipped one.

The benchmark table itself does not hand Qwen3.8-Max a clean sweep. Fable 5 leads on SWE-bench Pro, FrontierSWE, AndroidBench, CoWorkBench, JobBench, HLE, and HLE with tools, several of them by wide margins. GPT-5.6 tops Terminal Bench 2.1, DeepSWE 1.1, and GPQA Diamond. Qwen3.8-Max’s strongest columns are PaperBench, HealthBench, OSWorld-Verified, Parametric CAD Bench, and two multimodal robustness tests, Dense200 and VLMsAreBiased, where it posts double-digit leads over every rival listed.

For builders, the useful move now is narrow: QwenCloud already exposes an Anthropic-compatible API surface, which means tools like Claude Code, Codex, or Qoder CLI can already call this model today without waiting for the open weights. Hold judgment on the open-source claim until the Hugging Face and ModelScope uploads actually land next week, and treat the three autonomy showcases as claims to watch, not results to cite, until someone outside Qwen reruns them.

Qwen, the Alibaba-owned model team, published these claims and benchmark figures in its own Qwen3.8-Max release post on 3 August 2026.