Qwen3.8-Flash-Next: 125B Params, 6B Active, Outperforms with 1/3 Compute
Qwen team released the Qwen3.8-Flash-Next architecture report. It has 125B total parameters, 6B active, plus a 51B n-gram embedding table in host memory. With 1/3 active parameters, 1/3 training tokens, and about 1/9 training compute, it outperforms in 8 out of 14 benchmarks.