Alibaba's Qwen team has released Qwen3.8-2.4T, a 2.4 trillion-parameter Mixture-of-Experts model with approximately 95B active parameters per token, available in FP8 quantized format on HuggingFace. The release generates significant community interest on Hacker News with 416 points and 87 comments.
Background
Alibaba's Qwen series has been one of the most competitive open-weight LLM families in 2025-2026, regularly challenging proprietary models. This release continues the trend of scaling MoE architectures to massive parameter counts while maintaining efficiency through sparse activation.
- Source
- Hacker News (RSS)
- Published
- Aug 12, 2026 at 11:01 PM
- Score
- 7.0 / 10