logic65/Qwen3.8-Whittle-MoE-27B-A17.8B
27B • Updated • 33k • 111
Whittle: models you can run
Compressed Qwen3.8-27B variants with weights published and tested. Start here.
-
logic65/Whittle-Next-26B-A3B
Text Generation • 27B • Updated • 70Note qwen4_exp format on stock llama.cpp: sigmoid GDN, hyper-connections, 1B n-gram memory, online-distilled from Qwen3.8-27B. GSM8K 87.0%, chats and stops. Precursor to the 27B.
logic65/Whittle-Next-27B-A3B
Text Generation • 28B • Updated • 111 • 1Note 27B: 26B + 2B n-gram table + HC streams unlocked; stock llama.cpp; see caveats