Cerebellum target: Mistral Large 3 (675B MoE, 41B active)

[deucebucket]

Model

• Model: mistralai/Mistral-Large-3
• Total params: 675B
• Active params: 41B
• Architecture: MoE with routed experts

Why Cerebellum

Massive MoE with high community demand for local deployment. Standard quants still require 48+ GB VRAM. Cerebellum's ablation-guided mixed-precision could find significant savings — expert weights in large MoE models show high variance in quantization sensitivity.

What's needed

Too large for single 24 GB GPU ablation. Need someone with multi-GPU setup to run group and layer ablation experiments.

Steps:

1. Get BF16 GGUF + imatrix (bartowski usually has these)
2. Group ablation — crush each tensor category to Q2_K, measure PPL
3. Layer ablation on sacred/demotable groups
4. Share PPL logs here

Happy to build the final override file and quant from your data.