Qwen3.8-Max is live now. Qwen released the model on August 3, 2026, and called it a new bar for coding and cowork. Qwen3.8-Max is a 2.4 trillion-parameter Mixture-of-Experts model with 95 billion active parameters. It’s now widely accessible, and a full open-weight release is expected within the week.
What’s New
- Qwen3.8-Max supports a context window of up to 1 million tokens. You can access it now through Alibaba Cloud’s Model Studio API.
- The model benchmarks against frontier systems including Claude Opus 4.8, Claude Fable 5, and OpenAI’s GPT-5.6 Sol. Qwen evaluated it using established harnesses such as Terminus 2 and Claude Code. Tests included Terminal-Bench and SWE-bench Pro.
- Qwen reports real gains in reasoning, coding, agent capabilities, and multimodal understanding over its predecessor. The model also runs at lower compute cost.
- A coding harness built around Qwen3.8-Max has evolved through continuous autonomous operation. Qwen reports 265 commits and 127 pull requests after roughly 16 days of fully autonomous work. The harness gained new features along the way, including goal tracking, session resume, dynamic workflow management, and session replay.
- Qwen also launched QwenWork, an all-in-one workplace AI platform for individuals and businesses. It entered public beta the same day with both web and desktop apps.
- A full open-weight release is expected within the week. That would make Qwen3.8-Max one of the largest open-weight models to date, once the weights actually ship.
Why It Matters
The benchmark methodology here deserves as much attention as the scores themselves. Qwen compared Qwen3.8-Max directly against Opus 4.8, Fable 5, and GPT-5.6 Sol using established evaluation harnesses like Terminus 2 and Claude Code. That makes the reported results easier to compare than broad marketing claims, though independent benchmarking after release will still matter.
The autonomous development detail matters too. A coding harness gained hundreds of commits and pull requests through 16 days of unsupervised operation. That’s a real demonstration of the long-horizon, self-directed work Qwen is positioning this model around, not just a claim about what it can theoretically do.
Should You Care?
If you’re a developer evaluating models for coding or agent work, test Qwen3.8-Max directly through the Model Studio API now. Revisit it once the open weights ship, since running it yourself removes any dependency on Alibaba’s own infrastructure and pricing.
If you don’t build with AI models directly, QwenWork is the more relevant piece here. It’s Alibaba’s entry into workplace AI platforms already being built by Tencent and Moonshot AI. “AI at work” is becoming its own product category, not just a feature bolted onto a chat app.