XNews

Qwen3.8-Flash-Next: 125B Total, 6B Active, Outperforms 397B Model on 8/14 Benchmarks

Qwen released Qwen3.8-Flash-Next: 125B total params, 6B active, 51B n-gram embedding in host memory. Achieved 1/3 active params, 1/3 training tokens, ~1/9 compute of 397B-A17B flagship, surpassing it on 8 of 14 benchmarks.

Source posts