Frontier labs do not usually publish their best model. That is the business.
Alibaba released Qwen 3.8-Max on August 3 and said the open weights would follow the week of August 10. That week is now.
It would be the first time Alibaba has open-sourced a Max-class model, and it is the part of this launch that actually matters.
The Specifications
Qwen 3.8-Max is a sparse mixture-of-experts model with 2.4 trillion total parameters. Roughly 95 billion are active per token, about 4% of the total.
That active figure is the one that drives serving cost and latency. A 2.4 trillion parameter headline does not mean 2.4 trillion parameters run on every request.
The context window is one million tokens, with output capped around 131,000 tokens per response. It accepts text and visual input.
Coverage differs on the full modality list. Video, documents and speech have all been reported. Alibaba has not published a complete spec sheet, so treat the longer lists as unconfirmed.
A second checkpoint, Qwen 3.8-27B, is also going open-weights. For most developers that smaller model is the practically useful one, because a 2.4 trillion parameter model is not something you run at home.
| Attribute | Qwen 3.8-Max |
| Total parameters | 2.4 trillion |
| Active per token | about 95 billion |
| Architecture | Sparse mixture-of-experts |
| Context window | 1 million tokens |
| Launched | August 3, 2026 |
| Open weights | Scheduled week of August 10, 2026 |
| API pricing | $2 per million input, $6 per million output |
The Pricing Is the Strategy
Alibaba set API pricing at $2 per million input tokens and $6 per million output tokens. That is direct parity with OpenAI’s GPT-5.6.
This breaks the pattern people expect from Chinese labs. The usual playbook is to undercut heavily, and DeepSeek still does, at a fraction of those rates.
Pricing at parity is a claim about capability rather than cost. The message is that this model competes on quality, not on being the cheap option.
Then the weights get released anyway, which undercuts Alibaba’s own API. That combination only makes sense if the goal is distribution and standard-setting rather than near-term margin.
What the Benchmarks Say, and Who Is Saying It
Alibaba’s self-reported numbers include 86.6 on Terminal-Bench 2.1, 93.0 on PaperBench, and 67.7 on SWE-bench Pro. Another reported figure puts it at 86.1 on OSWorld-Verified.
Every one of those is vendor-reported. Independent verification across the full set remained limited as of the second week of August.
Third-party signal is more mixed than the vendor numbers suggest. On the crowdsourced Arena.AI leaderboard, Qwen 3.8-Max became the highest-ranked Chinese model for text tasks while still trailing several Anthropic models. For vision it ranked second globally.
Coverage also disagrees on the comparison set, with some outlets benchmarking against Claude Fable 5 and others against Claude Opus 4.8. Those are different models, so a claim of beating one is not a claim of beating the other.
The pattern is consistent, though. Leading on some agentic and coding tests, close but behind on general leaderboards.
The Agentic Pitch
The launch centers on long-running autonomous work rather than chat.
In one company-run test, Alibaba says the model spent about 16 days building and maintaining a command-line tool project, producing 265 commits, 127 pull requests and 151 issues through a loop of implementation, testing and repair.
In another, it says the model recreated a research pipeline, completed 33 GPU training rounds, and produced a method scoring 2.7 points above the original paper’s approach on a benchmark.
These are internal tests reported by the vendor with no external replication. They are interesting as a statement of what the model is built for, not as evidence it works.
Alibaba is also shipping API compatibility with both OpenAI and Anthropic interfaces, which lets developers point existing agent frameworks at it with minimal changes.
That compatibility layer is quietly the most aggressive part of the release. It reduces switching cost to almost nothing.
The Competitive Picture
Qwen 3.8-Max is the second-largest publicly known model. Moonshot AI’s Kimi K3 is larger at 2.8 trillion parameters and also shipped as open weights.
Two Chinese labs releasing frontier-scale open models within weeks of each other is not a coincidence. Release cycles among Chinese labs have compressed sharply.
The model was previewed on July 19 at the World AI Conference in Shanghai before the August 3 launch. Reporting indicates the open release had slipped previously as compute was redirected.
The Part Enterprise Buyers Should Read Twice
Open weights and the hosted API are two different products with two different risk profiles.
If you download the weights and run them on your own infrastructure, your data stays with you. That is the entire appeal for regulated industries.
If you send prompts to the hosted API, that traffic is processed under Chinese data jurisdiction, including the Cybersecurity Law and the Data Security Law.
For anything sensitive, that distinction is the whole decision. The same model name sits on both sides of it.
Sovereign and self-hosted AI infrastructure is becoming a live procurement question rather than a theoretical one, a theme running through World Mobile’s move to extend its node network into sovereign AI infrastructure.
Why This Matters Beyond AI
Open frontier weights change who can build autonomous systems without asking permission from a lab.
That is directly relevant to the agent economy, a shift Optimisus examined in why AI agents may become crypto’s next major user base.
It also lowers the floor for exchange-level and consumer-level AI tooling, the category that produced things like KuCoin’s crypto-native assistant rollout.
What to Watch This Week
The weights themselves. Until files appear on Hugging Face and ModelScope, this remains an announcement rather than a release, and Alibaba’s open timeline has slipped before.
Check the license terms when they land. Open weights and open source are not the same thing, and usage restrictions vary widely between releases.
Then wait for independent benchmarks. Vendor scores set expectations. They do not settle anything.
Sources
- South China Morning Post, Alibaba’s Qwen3.8-Max made widely accessible ahead of open-weights release — https://www.scmp.com/tech/article/3362738/alibabas-ai-model-qwen38-max-made-widely-accessible-ahead-open-weights-release
- The Daily Star, Alibaba releases Qwen 3.8-Max, its largest AI model yet, August 3, 2026 — https://www.thedailystar.net/news/tech-startup/news/alibaba-releases-qwen-38-max-its-largest-ai-model-yet-4238986
- MarkTechPost, Alibaba Qwen releases Qwen3.8-Max, a 2.4 trillion parameter MoE model — https://www.marktechpost.com/2026/08/03/alibaba-qwen-releases-qwen3-8-max/
- TestingCatalog, Qwen released Qwen3.8-Max with open weights coming soon — https://www.testingcatalog.com/qwen-released-qwen3-8-max-with-open-weights-coming-soon/
- Yotta Labs, Qwen 3.8-Max release date, specs and access — https://www.yottalabs.ai/post/qwen-3-8-max-release-date-specs-how-to-access-2026
- Forkast, Alibaba matches US closed-model pricing on eve of open-weights drop — https://forkast.news/qwen3-8-max-the-capability-war-begins-alibaba-matches-us-closed-model-pricing-on-eve-of-open-weights-drop/
Optimisus covers crypto and technology news for readers who want the detail behind the headline.
