Coding-focused Cerebellum builds (per-group HumanEval ablation).
AI & ML interests
Deep experimentation on ablation and quanting. Goal: trim the fat on models. Question: Is all data, good data, or, is some just noise?
Recent Activity
Weights under ~14 GB so a 16 GB GPU runs them fully loaded with room for context. Measured footprints on each card.
-
deucebucket/Qwen3.6-35B-A3B-Cerebellum-GGUF
Text Generation • 35B • Updated • 6.48k • 12 -
deucebucket/Qwen3.6-35B-A3B-Heretic-Cerebellum-GGUF
Image-Text-to-Text • 35B • Updated • 2.6k • 7 -
deucebucket/Gemma-4-26B-A4B-it-Cerebellum-v6-GGUF
Text Generation • 25B • Updated • 211 • 10 -
deucebucket/Gemma-4-26B-A4B-it-Heretic-Cerebellum-GGUF
Text Generation • 25B • Updated • 489 • 7
-
deucebucket/Gemma-4-26B-A4B-it-Cerebellum-v6-GGUF
Text Generation • 25B • Updated • 211 • 10 -
deucebucket/Gemma-4-E2B-it-Cerebellum-v2-GGUF
Image-Text-to-Text • 5B • Updated • 98 • 3 -
deucebucket/Gemma-4-E4B-it-Cerebellum-v2-GGUF
Image-Text-to-Text • 8B • Updated • 98 • 3 -
deucebucket/Gemma-4-26B-A4B-it-Cerebellum-GGUF
Text Generation • 25B • Updated • 128 • 4
Cerebellum mixed-precision quants of Heretic (abliterated) variants. Same recipes as the stock releases, transferred verbatim.
Cerebellum builds whose weights leave context headroom on 8 GB GPUs. Measured footprints are on each card.
The big ones. Includes the 122B with its measured expert-offload recipe (about 18 GB VRAM plus 34 GB RAM).
-
deucebucket/Qwen3.5-122B-A10B-Cerebellum-GGUF
Text Generation • 122B • Updated • 92 • 5 -
deucebucket/Qwen3-32B-Cerebellum-GGUF
33B • Updated • 37 • 1 -
deucebucket/Granite-4.0-H-Small-Cerebellum-GGUF
Text Generation • 32B • Updated • 24 • 1 -
deucebucket/KAT-Coder-V2.5-Dev-Cerebellum-GGUF
Text Generation • 35B • Updated • 301 • 3
Cerebellum mixed-precision quants of Heretic (abliterated) variants. Same recipes as the stock releases, transferred verbatim.
Coding-focused Cerebellum builds (per-group HumanEval ablation).
Cerebellum builds whose weights leave context headroom on 8 GB GPUs. Measured footprints are on each card.
Weights under ~14 GB so a 16 GB GPU runs them fully loaded with room for context. Measured footprints on each card.
-
deucebucket/Qwen3.6-35B-A3B-Cerebellum-GGUF
Text Generation • 35B • Updated • 6.48k • 12 -
deucebucket/Qwen3.6-35B-A3B-Heretic-Cerebellum-GGUF
Image-Text-to-Text • 35B • Updated • 2.6k • 7 -
deucebucket/Gemma-4-26B-A4B-it-Cerebellum-v6-GGUF
Text Generation • 25B • Updated • 211 • 10 -
deucebucket/Gemma-4-26B-A4B-it-Heretic-Cerebellum-GGUF
Text Generation • 25B • Updated • 489 • 7
The big ones. Includes the 122B with its measured expert-offload recipe (about 18 GB VRAM plus 34 GB RAM).
-
deucebucket/Qwen3.5-122B-A10B-Cerebellum-GGUF
Text Generation • 122B • Updated • 92 • 5 -
deucebucket/Qwen3-32B-Cerebellum-GGUF
33B • Updated • 37 • 1 -
deucebucket/Granite-4.0-H-Small-Cerebellum-GGUF
Text Generation • 32B • Updated • 24 • 1 -
deucebucket/KAT-Coder-V2.5-Dev-Cerebellum-GGUF
Text Generation • 35B • Updated • 301 • 3
-
deucebucket/Gemma-4-26B-A4B-it-Cerebellum-v6-GGUF
Text Generation • 25B • Updated • 211 • 10 -
deucebucket/Gemma-4-E2B-it-Cerebellum-v2-GGUF
Image-Text-to-Text • 5B • Updated • 98 • 3 -
deucebucket/Gemma-4-E4B-it-Cerebellum-v2-GGUF
Image-Text-to-Text • 8B • Updated • 98 • 3 -
deucebucket/Gemma-4-26B-A4B-it-Cerebellum-GGUF
Text Generation • 25B • Updated • 128 • 4
Cerebellum mixed-precision quants of Heretic (abliterated) variants. Same recipes as the stock releases, transferred verbatim.
Cerebellum mixed-precision quants of Heretic (abliterated) variants. Same recipes as the stock releases, transferred verbatim.