Qwen3 235B MoE Model Weights Released on HuggingFace and ModelScope

Alibaba's Qwen team has released weights for a large Mixture-of-Experts model, Qwen3, publicly on HuggingFace and ModelScope.

A new large-scale AI model from Alibaba's Qwen team has appeared on HuggingFace and ModelScope under the name Qwen3.8-2.4T-A95B-FP8, suggesting a Mixture-of-Experts architecture with roughly 2.4 trillion total parameters and 95 billion active parameters, stored in FP8 precision. The model weights are publicly accessible via HuggingFace at the repository Qwen/Qwen3.8-2.4T-A95B-FP8, and a corresponding release has been noted on ModelScope, Alibaba's own model-sharing platform. Beyond the repository link and model name, the sources provide no additional technical documentation, benchmarks, license terms, or release notes. The naming convention follows Qwen's established pattern for MoE models, where total and active parameter counts are distinguished — but all specifics should be treated as inferred from the model name alone until official documentation is available.

Why it matters

If the parameter counts implied by the name are accurate, this would be one of the largest openly accessible MoE models released to the public, potentially relevant to researchers and developers building on frontier open-weight AI. However, without confirmed documentation, the model's actual capabilities, licensing, and intended use remain unclear.

What's next

Researchers and the AI community will likely examine the model weights and publish independent evaluations; official technical documentation or a blog post from the Qwen team would clarify confirmed specifications.

Key facts

• Model name: Qwen3.8-2.4T-A95B-FP8, hosted at huggingface.co/Qwen/Qwen3.8-2.4T-A95B-FP8
• Name implies approximately 2.4 trillion total parameters and 95 billion active parameters in a Mixture-of-Experts architecture
• Weights are stored in FP8 (8-bit floating point) precision
• Released simultaneously on HuggingFace and ModelScope (Alibaba's model platform)
• No technical report, benchmark results, or license terms were included in the available source text

Bias & framing notes

All three sources provide virtually no body text — only a model name and a single URL. No independent reporting, official press release, or technical documentation is present. All substantive details in this summary are inferred from the model's naming convention and should be treated as unverified until confirmed by Qwen's official release materials.

NewsClear — neutral news & congressional tracking · Bill of the Week