mannix/ gemma4-98e-v4:Q8_0

89 3 days ago

Pruned to 98 experts gemma-4 a4b 26b v4

tools thinking 
ollama run mannix/gemma4-98e-v4:Q8_0

Details

3 days ago

59ec01dfea3c · 21GB ·

model
gemma4
·
19.9B
·
Q8_0
template
{{- if or .System .Tools }}<bos><|turn>system {{ if .System }}{{ .System }} {{ end }}{{- if .Tools }
params
{ "num_ctx": 256000, "repeat_last_n": 256, "repeat_penalty": 1.15, "stop": [

Readme

The gemma-4-A4B-98e-v4 is pruned specifically to keep general knowledge as wide as possible, unlike the v3 aimed specifically at keeping intact reasoning. The token usage is similar as the original 128e version, lower than v3 which needs 1.7x.

  HumanEval-chat token usage (164 problems × max=3072)

  ┌──────────────┬─────┬─────┬─────┬─────┬──────┬─────┐
  │   variant    │ min │ p10 │ p50 │ p90 │ max  │ avg │
  ├──────────────┼─────┼─────┼─────┼─────┼──────┼─────┤
  │ 128e @3072   │  35 │ 125 │ 314 │ 589 │  917 │ 334 │
  ├──────────────┼─────┼─────┼─────┼─────┼──────┼─────┤
  │ 98e-v4       │  35 │ 114 │ 304 │ 648 │  895 │ 340 │
  ├──────────────┼─────┼─────┼─────┼─────┼──────┼─────┤
  │ 98e-v3 @3072 │  35 │ 206 │ 490 │ 897 │ 1013 │ 512 │
  └──────────────┴─────┴─────┴─────┴─────┴──────┴─────┘

Template fixed for tools usage

Model on HF:

https://huggingface.co/ManniX-ITA/gemma-4-A4B-98e-v4-it

Full GGUF:

https://huggingface.co/ManniX-ITA/gemma-4-A4B-98e-v4-it-GGUF