Alibaba’s model team Qwen officially released its large language model Qwen3.8-Max on August 3, 20261. The model has 2.4 trillion total parameters with 95 billion active, and the company’s blog says this is the first time it will open-source the weights of a Qwen-Max-class model, with the weights due out the following week1.
When we covered the July 19 preview announcement, benchmarks, licensing, and a release date were all unpublished. Pricing and benchmarks have now been filled in, and the weight release has at least a rough timeframe.
Pricing and access
The model is served through QwenCloud and called via API1. The pricing page lists $2 per million input tokens, $6 per million output tokens, and a 1-million-token context window2. The official blog’s setup examples include a "context_window": 1000000 setting and a step that sets ANTHROPIC_AUTH_TOKEN1, so existing tooling can be pointed at an Anthropic-compatible endpoint to try it.
The architecture builds on Qwen 3.5, scaled up to 2.4 trillion parameters1. Of those, 95 billion are active. The company has also said on its official X account that a smaller Qwen3.8-27B will go open-weights the same week3.
Licensing terms, however, remain unpublished. The blog offers only the phrase “open weights” and the timing “next week” — nothing that would let a company judge whether commercial use is permitted. As with K-EXAONE 2.0, which was released under Apache 2.0, adoption decisions only become possible once the terms are explicit, so that part has to wait.
The official benchmark table cuts both ways
Alibaba published a comparison table against Opus 4.8, Fable 5, GPT-5.6 Sol (max), and its own previous-generation Qwen3.7-Max1. The main coding-related figures are below — all measured and published by Alibaba itself1.
| Benchmark | Opus 4.8 | Fable 5 | GPT-5.6 Sol (max) | 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 |
| AndroidBench | 69.8 | 84.5 | 74.0 | 56.5 | 75.1 |
The gains over Qwen3.7-Max are large across the board — DeepSWE 1.1 goes from 21.6 to 56.61. Against rival models, though, the direction changes by benchmark. Qwen3.8-Max posts 93.0 on PaperBench, the highest figure in the table, but scores 67.7 on SWE-bench Pro against Fable 5’s 80.0, and 56.6 on DeepSWE 1.1 against Fable 5’s 70.0 and GPT-5.6 Sol (max)‘s 73.0 — those losses are printed in the company’s own table1. The “second only to Fable 5” framing Alibaba used in July is not uniformly borne out by these numbers.
The figures also need reading with care. Per the footnotes, Qwen3.8-Max was measured on Terminal Bench 2.1 with Claude Code at avg@10, while for other models Alibaba reports the best published score across harnesses1. For SWE-bench Pro, it says problematic tasks were corrected and all baselines re-evaluated on the refined benchmark1. That the same benchmark name can move with the execution setup was demonstrated only days ago when OpenAI reported that two API settings roughly tripled its ARC-AGI-3 score. These numbers are conditions-attached, not a flat ranking.
Several entries in the table — QwenSWEBench and QwenQoderBench among them — are noted in the footnotes as Alibaba’s own in-house benchmarks1. Third-party verification has yet to come.
265 commits over 16 days of autonomous operation
What Alibaba foregrounded in this release was not one-shot code generation but long-running autonomous operation. It had the model build a project called oh-my-cli from scratch and, over a 10-plus-day autonomous coding run, construct a harness that improves itself1. Requirements are normalized into GitHub Issues, an agent claims them through a state machine, and after implementation, end-to-end tests and checks run before the PR is merged1.
The company says that as of July 30, 2026, roughly 16 days of fully autonomous operation had produced 265 commits, 127 PRs, and 151 issues in the repository1. The project trace is available in a public repository1. In a separate task, it says it asked the model to reproduce an experiment from a paper on arXiv in code and then improve on it1.
Whether running a model that way is practical does not come down to the model alone. Much of it rests on the operational design — how the agent is started, where it is stopped, and what permissions it holds. Providers are still building out the stopping mechanisms, as when Gemini API’s managed agents gained hooks that can block tool calls.
Trillion-scale weight releases keep coming
Weight releases from large models have arrived in quick succession over the past few weeks. Moonshot AI published the full weights for Kimi K3 in late July, and LG AI Research released the 750B K-EXAONE 2.0 under Apache 2.0. If Qwen3.8-Max’s weights land as promised, the set of self-hostable models claiming frontier-tier capability grows again.
That said, few organizations can run a 2.4-trillion-parameter model themselves. The more practical development is the published pricing. At $2 in and $6 out with a 1-million-token context, there is now enough to model whether it could displace existing API spend. On how open weights should be treated in the first place, Anthropic has said it is not arguing for a ban while narrowing its regulatory asks to three points — the debate is shifting from whether to publish toward the conditions attached.
For an evaluation, the quickest route is to swap in the Anthropic-compatible endpoint on a small slice of work and measure it on your own tasks. A real adoption decision can wait until the license terms are published and third-party benchmarks exist.
Sources
- Qwen3.8-Max: A New Bar for Coding and Cowork - Official announcement by the Qwen team (August 3, 2026)
- QwenCloud - Official pricing and specification listing
- Alibaba Qwen official X post - Announcement that Qwen3.8-27B will go open-weights