Qwen prepares the Flash-Next model: 125 billion parameters with 6 billion active

A page for the Qwen3.8-Flash-Next model has appeared on ModelScope, as reported on Hacker News (122 points, 47 comments). Little can be gleaned from the sparse data, but even that little is intriguing.

Judging by the 125B a6B designation, this is a Mixture-of-Experts architecture: the total number of parameters is around 125 billion, while each individual pass engages only about 6 billion. This approach makes it possible to maintain a high model capacity while noticeably saving computational resources during inference — and it is speed that the word Flash in the name hints at, too.

The Next suffix and the mention of a release “tomorrow” indicate that the weights and accompanying materials should be published in the near future. At the time the page appeared, there was still no detailed documentation, benchmarks or official announcement — only the model's name and its basic specifications were available.

For now this is more a reason to keep watching than to draw conclusions: without a model card, a license and test results, it is premature to judge how Flash-Next will differ from previous Qwen releases. We await the official details.

Source: Hacker News