Spaces:
Running on Zero
Apply for a GPU community grant: Personal project
Cerebellum is an open-source method for compressing large language models using per-tensor sensitivity ablation. Instead of applying uniform quantization, we measure how each individual tensor responds to precision changes — testing hundreds of configurations — then build a custom mixed-precision GGUF that preserves quality where it matters and crushes aggressively everywhere else.
Our Qwen 3.6 27B Cerebellum v4 achieves 75% HumanEval · 95% ARC · 91% HellaSwag · 77% MMLU-Redux at just 12 GB — 181 per-tensor overrides tuned through systematic ablation.
GPU time is the bottleneck — each tensor ablation requires a full quantize + perplexity evaluation cycle. With 658 tensors per model and only 14% tested so far, systematic coverage across model families would let us build a transferable "quantization sensitivity atlas" — mapping which tensor types and layer positions are universally sensitive vs. crushable across architectures.
All results, override files, and methodology are published openly. The goal: make 27B+ models run on consumer GPUs with minimal quality loss.