gghfexp/lmheads

Original output.weight tensors punched out of GGUF models

Naming (loosely):

original_model_original_precision_weighttype.f32

Notes

Can swap K2-Thinking <-> K2-Instruct Can swap K2.5, K2.6, K2.7 interchangably Can swap GLM-4.6 <-> GLM-4.7

Cannot swap [K2-OG/K2-I/K2-T] <-> [K2.5/K2.6/K2.7] (incoherent)

Official Model Heads (from BF16)

kimi_k25_bf16_ow.f32 tensor output.weight type=30 ne=[7168,163840] k=1174405120

kimi_k26_bf16_ow.f32 tensor output.weight type=30 ne=[7168,163840] k=1174405120

kimi_k27_coder_bf16_ow.f32 tensor output.weight type=30 ne=[7168,163840] k=1174405120

kimi_k2_instruct_bf16_ow.f32 tensor output.weight type=30 ne=[7168,163840] k=1174405120

kimi_k2_thinking_bf16_ow.f32 tensor output.weight type=30 ne=[7168,163840] k=1174405120

Heads ripped off community finetunes

gryphe_gemma4-31b_bf16_ow.f32

Runtime Hot-Swap

curl --request POST \
  --url http://localhost:8080/style-head \
  --header 'content-type: application/json' \
  --data '{
  "path": "/workspace/lmheads/kimi-k2_thinking/kimi_k2_thinking_bf16_ow.f32"
}'

curl --request POST \
  --url http://localhost:8080/style-head \
  --header 'content-type: application/json' \
  --data '{
  "path": "/workspace/lmheads/gryphe_gemma4-31b/gryphe_gemma4-31b_out.f32"
}'