logic65/Qwen3.8-Whittle-16B
Text Generation • 16B • Updated • 6.12k • 2
Qwen3.8 Whittle family
A 27B whittled on two 8GB GPUs: measured cuts, heals, and restorations. 48L = best untrained, Whittle-16B = best overall.
-
-
logic65/Qwen3.8-Whittle-48L-cut
Text Generation • Updated • 1 -
logic65/Qwen3.8-Whittle-44L-cut
Text Generation • Updated • 1 -
logic65/Qwen3.8-p44w75-16.8B-unrepaired
Text Generation • 16B • Updated • 7.46k • 2 -
logic65/Qwen3.8-Whittle-v40-16.8B
Text Generation • 18B • Updated • 414 • 1 -
logic65/Qwen3.8-Whittle-w50-18.3B-unrepaired
Text Generation • 18B • Updated • 1.14k • 1 -
logic65/Qwen3.8-Whittle-tri-14.7B
Text Generation • 15B • Updated • 2.79k • 1
logic65/Qwen3.8-Whittle-MoE-27B-A17.8B
27B • Updated • 27.7k • 107Note Best of the family at 32/39. Sparse MoE carved from the dense 27B: 5120 shared + 64 experts of 192, top-16, so 17.8B of 26.9B runs per token. English only, 0/3 multilingual.