Repository navigation

Add the IQ1_M, IQ2_XXS and IQ2_S weight-only quantization formats - #2505

Closed

cjluo-nv wants to merge 6 commits into

mainfrom

chenjiel/iq-1bit-2bit-formats

Closed

Add the IQ1_M, IQ2_XXS and IQ2_S weight-only quantization formats#2505
cjluo-nv wants to merge 6 commits into
mainfrom
chenjiel/iq-1bit-2bit-formats

Commits on Sep 22, 2026
- Add the IQ1_M, IQ2_XXS and IQ2_S weight-only formats
  
  cjluo-nv
  and
  claude
  committed
  0ec9d71
- Note the new IQ formats in the changelog
  
  cjluo-nv
  and
  claude
  committed
  21eb1af
- Use the IQ1_M scale anchor rather than IQ1_S's
  
  cjluo-nv
  and
  claude
  committed
  0243784
- Add CUDA encoders for IQ1_M, IQ2_XXS and IQ2_S
  
  cjluo-nv
  and
  claude
  committed
  178efd3
- Bring every GGML IQ format to the same tests and surface
  
  cjluo-nv
  and
  claude
  committed
  0892bda
- Drop the torch-only caveat from the IQ changelog entry
  
  cjluo-nv
  and
  claude
  committed
  029821f