Skip to content

Cerebellum target: DeepSeek-V4 (1.6T MoE, 49B active) #4

Description

@deucebucket

Model

  • Model: deepseek-ai/DeepSeek-V4
  • Total params: 1,600B
  • Active params: 49B
  • Architecture: MoE (ships with FP4/FP8 mixed-precision by default)

Why Cerebellum

DeepSeek-V4 already ships with its own mixed-precision scheme (FP4 for expert weights, FP8 elsewhere). But that's a uniform policy — Cerebellum could improve on it by testing which specific expert layers benefit from higher or lower precision. The model is ~860 GB even with default mixed-precision.

High community demand for better local deployment options.

What's needed

Way too large for consumer hardware ablation. Need access to serious iron (8x A100/H100 or equivalent).

If you have the hardware:

  1. Group ablation — test each tensor category at different precisions
  2. Layer ablation on groups that show variance
  3. Share PPL logs — we build the override file

Even partial data (just group-level ablation) would tell us if the method works at this scale.

Metadata

Metadata

Assignees

No one assigned

    Labels

    help wantedExtra attention is needed

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions