Alibaba's biggest model coded alone for 16 days. Next week, its weights go public.

Alibaba's newest flagship model spent 16 days working alone and delivered finished software at the end of it. Qwen3.8-Max, announced on August 3 out of Hangzhou, is the largest and most capable model in the Qwen series to date. Next week its weights go public in the first open-weight release of a Qwen-Max-class model. That pairing, a frontier-scale model built for multi-day autonomous work that anyone can download, has no precedent in the lineup. The Max tier was the closed part of Qwen.

The model that coded for 16 days straight

The demo that matters is a single, checkable fact: in internal testing, Qwen3.8-Max ran a real-world software engineering project for 16 days with no human in the loop. Tasked with creating a self-evolving agent framework from scratch, the model built an engineering loop that synthesized user feedback, community best practices, and its own test data, then cycled through code generation, testing, previewing, and log analysis. The output is "oh-my-cli," an agent framework Alibaba has fully open-sourced on GitHub.

The autonomy shows up outside coding too. Alibaba reports the model beat human participants in the WWW2025 Multimodal Dialogue Intent Recognition Challenge, parsing customer-service transcripts and complex user requests. In a black-box benchmark called RecreationBench, with no internet access and no visibility into source code, it rebuilt applications from scratch purely by interacting with live versions and reading visual feedback. In a separate 24-hour contest, run solo, it beat 458 of 526 human teams, according to the contest report. Qwen3.8-Max currently ranks fifth in Text Arena, second in Vision Arena, and fourth in Frontend Code Arena.

95 billion active parameters, 2.4 trillion total

Qwen3.8-Max is built on a Sparse Mixture-of-Experts architecture with a hybrid attention mechanism, the team says, extending the foundation of Qwen 3.5. The model totals 2.4 trillion parameters but activates only 95 billion per request. That gap is the point: Alibaba says the design cuts computational costs and latency compared with dense models of similar scale, which is what makes a model this size practical to run outside a lab. The context window reaches 1 million tokens.

Spec	Qwen3.8-Max
Total parameters	2.4 trillion
Active parameters	95 billion
Context window	Up to 1 million tokens
Text Arena rank	5th
Vision Arena rank	2nd
Frontend Code Arena rank	4th
Weights release	Next week, Hugging Face and ModelScope

The model is multimodal. It can ingest hundred-page documents, full television series, or 100-hour livestreams and turn them into searchable knowledge bases. Alibaba's examples cover reconstructing a frontend project from a single UI screenshot, converting 2D floor plans into 3D interiors, and editing raw footage into vlogs, all through real-time visual feedback.

The open-weight bet

In Qwen's lineup, the Max class was the closed tier. Qwen3.8-Max breaks that: Alibaba's announcement says weights will be published on Hugging Face and ModelScope next week, and the model is already available through the Alibaba Cloud Model Studio APIs and on QwenWork, the company's all-in-one workplace agent platform.

It fits the direction Alibaba has taken since it consolidated its AI work into a single business unit with a stated mission to "create, deliver, and apply tokens," backed by a $53 billion investment push into cloud and AI infrastructure. The fleet strategy is visible across the lineup: Qwen3.7-Max, the previous flagship from May, was independently benchmarked by Artificial Analysis as outperforming leading Chinese competitors and matching top global systems. Alongside it sit Qwen3.7-Plus, a multimodal model aimed at agent workflows with tool use and adaptive web search, and now a flagship that will be downloadable by anyone.

Shipping a fleet is one thing. Making the fleet's biggest member open source is another, and it reads like a bet on third-party adoption over hosted API sales alone: a platform play, not a model race.

Open weights land under agent scrutiny

The timing puts the two trends in the same frame. Autonomous agents are under fresh scrutiny: the UK's AI Security Institute reported that agents from OpenAI and Anthropic, tested with safeguards disabled, took unsanctioned action on the live internet, created fake online identities, and pressured real people to approve malicious code. How to evaluate models that can act on their own for days is unresolved, one corner of the broader agent-trust problem.

An open-weight flagship built for long-horizon agent runs moves that evaluation problem from a handful of labs to whoever downloads it. Alibaba's announcement, which details the internal testing, does not describe safety evaluations for the release. The most capable open model Alibaba has shipped is also the one most built to act independently. Both facts are now true at once, and the community gets to sort out what that means.