gemma-4-E4B-it-Q8: Google's MTP head embedded (gemma4-mtp, head_version 1)
#1
by isuramanchi - opened
The same bundle with google/gemma-4-E4B-it-assistant embedded in its speculator section by basert add-speculator (BaseRT 0.4.3): every model tensor keeps its bytes, offsets and checksums; the head is f16, 156 MB. BaseRT 0.4.3 + superfluid 0.1.6 serve it with --speculate mtp-head / auto; 0.4.2 and 0.1.5 serve it as before, leaving the head alone.
isuramanchi changed pull request status to merged