Skip to content

An editorial team and a community building with AI

vibecoding.incJoin Discord ↗
← All articles

News · Alibaba (Qwen) · 2 min read

Event date: · Published:

Qwen publishes open weights for Qwen3.8-2.4T-A95B, a 2.4T-parameter Qwen-Max-class model

On August 12, 2026, Qwen published Qwen3.8-2.4T-A95B on Hugging Face Hub and ModelScope, which it described as the first open release of a Qwen-Max-class model, with 2.4T total and 95B activated parameters.

Archive entry: this covers an event from August 12, 2026, and we published it on November 8, 2026. The details come from sources dated at the time and may have changed since.

On August 12, 2026, Qwen published Qwen3.8-2.4T-A95B on Hugging Face Hub and ModelScope, according to the Qwen3.8 repository README. Qwen said that, for the first time, Qwen3.8 brought a Qwen-Max-class model to open release. Qwen3.8-Max was the official hosted version.

What changed

  • Architecture: a mixture-of-experts model with 2.4T total and 95B activated parameters, 92 layers and 512 experts (10 routed plus 1 shared), built on the Qwen3.5 architecture.
  • Context: 262,144 tokens natively, extensible to 1,010,000 tokens.
  • License: a custom qwen3.8-max license.
  • Serving: it worked with vLLM, SGLang and TokenSpeed, and an FP8 variant was available. It supported the reasoning_effort and preserve_thinking parameters.

What it meant for people building with AI

Teams that self-host could run a Qwen-Max-class model, and it worked with Qwen Code and with Anthropic-compatible clients through the QwenCloud API. The Alibaba Cloud Model Studio release list cited GPQA Diamond 92.6, PaperBench 93.0, OSWorld 86.1 and BabyVision 82.0, plus fourth place globally on CodeArena.

Access and the hosted version

Qwen3.8-Max, the official hosted version, added vision input, non-thinking support, a 1M default context and built-in tools. The open model was available through Qwen Studio (free), Qoder, QwenWork, the QwenCloud API (OpenAI and Anthropic compatible) and Qwen Code, and the weights were on the Hugging Face model page.

Sources

Written with AI assistance, strictly from the sources listed above. Before it went live, automated checks verified its dates, sources and every language edition.

Qwen publishes open weights for Qwen3.8-2.4T-A95B, a 2.4T-parameter Qwen-Max-class model | vibecoding.inc