pipenetwork/Qwen3.8-Flash-Next-MLX-8bit
Text Generation • 177B • Updated • 1.66k
Qwen3.8-Flash-Next MLX
Apple Silicon (MLX) builds of Qwen3.8-Flash-Next with a validated qwen4_exp runtime: github.com/PipeNetwork/qwen38-flash-next-mlx
Note 192 GB. Indistinguishable from bf16: ppl 4.4749 vs 4.4708, paired ratio 1.0009 [0.9997, 1.0021], better on 50% of windows.
pipenetwork/Qwen3.8-Flash-Next-MLX-6bit
Text Generation • 177B • Updated • 532 • 1Note 148 GB. Indistinguishable from bf16: ppl 4.4767, paired ratio 1.0013 [0.9997, 1.0029], better on 44% of windows.
pipenetwork/Qwen3.8-Flash-Next-MLX-mixed-4_8bit
Text Generation • 182B • Updated • 2.39k • 9Note 106 GB. ppl 4.5286 (+1.3%): 4-bit experts and n-gram tables, 8-bit for the 3% of non-expert weights. The build to use at this size.
pipenetwork/Qwen3.8-Flash-Next-MLX-4bit
Text Generation • 177B • Updated • 490 • 1Note 104 GB. ppl 5.3914 (+20.6%): the non-expert 3% at 4-bit costs 19 points of perplexity. Dominated by mixed-4_8bit for 2.4 GB more.