Alibaba has released Qwen 3.8-Max, its first open-weight model in the Max tier, scaling to 2.4 trillion total parameters with 95 billion active. The new architecture delivers top-tier performance on coding and long-horizon reasoning benchmarks, with full model weights dropping next week on Hugging Face and ModelScope.
Alibaba Drops Qwen 3.8-Max, Its First Open-Weight Model in the Max Tier
Alibaba has released Qwen 3.8-Max, and it’s bringing open weights along for the ride. The new model scales to 2.4 trillion total parameters with 95 billion active, making it the first Qwen-Max-class release to actually open its weights to the public.
It launched today, August 3, 2026, alongside a raft of benchmark numbers that put it in direct conversation with Claude Opus 4.8 and GPT-5.6 Sol. The catch is that the weights drop next week on Hugging Face and ModelScope. As of right now, you can only spin it up through the DashScope API. Alibaba had been teasing a Max-class open release since early summer. The initial rollout slipped twice as compute allocation got shuffled around to support the domestic cloud push. By the time Qwen 3.8-Max actually landed, the competitive landscape had shifted.
The Specs and the Benchmarks
The architecture is a sparse Mixture-of-Experts built on the Qwen 3.5 foundation. You get a one million token context window, 65,536 default output tokens (expandable to 131,072), and multimodal input for text and images. There’s also a reasoning effort toggle with three levels: xhigh, medium, and low. xhigh is the default.
On coding and reasoning benchmarks, Alibaba is pointing to some serious gains. PaperBench sits at 93.0, outpacing GPT-5.6 Sol and Claude Opus 4.8. IFBench lands at 82.8, well ahead of GPT-5.6 Sol’s 72.7. Terminal Bench 2.1 hits 86.6, and it wrapped a simulated 24-hour multimodal dialogue competition with 0.853 accuracy, beating 458 of the 526 human teams entering. Keep in mind that Alibaba evaluates on a custom QwenSWEBench alongside the standard SWE-bench Pro.
It’s a rather aggressive pricing strategy for a model of this size, though the API rates through QwenCloud undercut most of the closed competitors per token. Head here to the official QwenCloud dashboard for exact pricing tiers and token plans.
Long-Horizon Demos and Autonomous Workflows
Alibaba’s official blog walks through a few headline stunts. One involves letting the model run a chip design flow end-to-end. It cut gate count from 8,298 down to 678 and slashed physical layout area by 81%, all while hitting timing closure at 500 MHz. That’s not a toy demo.
Another stretched over 16 days, producing 265 commits and 127 pull requests for an open-source CLI project called oh-my-cli. The model iterated through its own code, tests, previews, and logs without human hand-holding. Then there’s a 365-day e-commerce simulation on real Taobao transaction data where it turned ¥100,000 into ¥416,252.
It also rebuilt running applications across Ubuntu, macOS, Windows, Android, and the web by watching them as black boxes. The team calls it a vision feedback loop. The model generates while observing and acts while reviewing. On top of that, Alibaba shipped a new plugin harness for image processing, multimodal memory, and dynamic resolution support.
API Access and the Open Weight Drop
API access starts today through DashScope, with endpoints in Beijing, Singapore, and Virginia. It speaks both OpenAI and Anthropic formats, which means you can point Claude Code or Codex at it without rewriting your prompts. Open weights arrive next week on Hugging Face and ModelScope. That’s the big shift. Previous Max-class releases stayed locked behind Alibaba’s cloud. You can now fine-tune, distill, or just poke around locally if your hardware can handle it.
However, at the same time, the API-only window is worth leveraging if you don’t want to manage your own GPU cluster. The model is heavy. Raw benchmark supremacy rarely tells the full story, though. Independent verification is welcome. Alibaba says the open release will fuel community fine-tuning and production deployments starting next month. Keep an eye on Hugging Face for the drop.
