Alibaba's AI team Qwen has released the model weights for Qwen3.8 openly. The release is another marker of how fast competition is moving on the open-weight side of the market.

The core model: 27 billion parameters

At the centre of the family sits Qwen3.8-27B, a multimodal dense model with 27 billion parameters. According to Qwen, it outperforms the larger Qwen3.7-Plus on coding and office tasks.

The team also claims improved agent capabilities, saying the model plans more independently and completes tasks more reliably. Those claims rest on the company's own measurements; independent verification is not yet available.

Context and modes

On the technical side the model offers:

  • A native context window of 262,000 tokens
  • Scaling to one million tokens using the YaRN method
  • Image and video processing alongside text — including diagrams, documents and multi-hour video
  • A flexible thinking mode, on by default and toggleable per query

Being able to switch thinking on and off per query is a practical detail: unnecessary reasoning on simple tasks costs both latency and money.

Licence and access

The weights ship under the Apache 2.0 licence — commercially permissive and relatively unrestricted, more open than the custom licences some rivals attach to their open-weight releases.

Qwen has also published weights for the much larger Qwen3.8-2.4T-A95B, built to operate at the Max level. Both models are available on Hugging Face and ModelScope.

A hosted version with a one-million-token context will soon be available through Qwen Cloud, Alibaba's AI service.

Why it matters

Open-weight models are the only realistic option for teams that cannot depend on closed APIs or that must keep data on their own infrastructure. A 27-billion-parameter model beating a larger one on specific tasks is another reminder that size alone is not decisive — and for anyone running models on their own hardware, that translates directly into cost.

Two models, two different needs

Shipping two models is not incidental. The 27-billion-parameter dense model fits on a single server or a capable workstation — the target for teams that want to run models on their own hardware. Qwen3.8-2.4T-A95B, built to operate at the Max level, is a different class entirely: the weights are open, but running it demands serious infrastructure.

Open weights do not automatically mean "anyone can run this." Downloadable weights allow inspection, fine-tuning and independent evaluation, but the cost of serving a model in production still sits in the hardware. That is precisely why a smaller model beating a larger one on specific tasks matters.

Whose measurements back the claims

The advantage on coding and office tasks, and the improved agent capabilities, are for now Qwen's own account. Independent benchmark results have not yet been published. This is one advantage of open-weight releases: the claims can be tested from the outside fairly quickly, because running the model does not require the company's permission.