Alibaba's Qwen team is preparing to release Qwen 3.8-Flash-Next on Wednesday, a 125-billion-parameter mixture-of-experts model that activates only 6 billion parameters per token. The release is described as an early preview of the next-generation Qwen 4 architecture, intended to help developers prepare for the full model family. The team has not yet published hard benchmark scores, and the weights were not live on ModelScope at the time of reporting, but the model is expected to be multimodal and open-weight.
The mixture-of-experts design allows the large model to run with the compute cost of a much smaller 6-billion-parameter system. That is significant because it puts near-frontier AI capability within reach of commodity hardware. Alibaba’s teaser follows a broader wave of open-weight releases from Chinese labs including DeepSeek and Moonshot, and a mysterious free model called Ox Alpha that recently beat Anthropic’s Fable on certain coding benchmarks.
Separately, researchers at Multiverse Computing published a method called Quantization-Aware Healing on the Hugging Face blog on August 25. They shrank OpenAI's open GPT-OSS model from 120 billion parameters to 60 billion and compressed its memory to 4-bit precision. In 7 of 9 tests, the smaller 4-bit 60-billion-parameter model outperformed the full-precision 60-billion-parameter version, although the original 120-billion-parameter model still won most rounds.
The key insight is that the compressed model was trained to imitate the original smart model rather than an intermediate lower-precision copy. The team released the resulting Hypernova-60B model as open weights on Hugging Face. However, the shrinking tool itself remains proprietary, and the method has so far only been tested on GPT-OSS, not on Llama, Qwen, or Mistral families.
