DeepSeek V4 Flash — REAP K128 (Mixed Precision)
REAP-pruned DeepSeek V4 Flash at K128 (128 routed experts per MoE layer). Prunes 50% of routed experts via Cerebras REAP (Router-weighted Expert Activation Pruning), preserving all attention, embeddings, shared experts, router, and MTP components.
At a Glance
| Base model | DeepSeek V4 Flash |
| Donor GGUF | antirez Layers37-42Q4KExperts (90.9 GiB) |
| Pruning method | REAP (Cerebras Research) |
| Routed experts | 128 per layer (down from 256) |
| Kept slots | 5,888 / 11,008 |
| Hash-preserved | Layers 0-2 (256 experts each) |
| Pruned | Layers 3-42 (128 experts each) |
| Format | ds4-compact-v1 GGUF |
| File size | 52.04 GiB |
| Mixed precision | Layers 3-36: Q2_K/IQ2_XXS · Layers 37-42: Q4_K |
Note: This variant preserves the donor's mixed quantization (Q4_K in layers 37-42, IQ2_XXS/Q2_K elsewhere). Requires a ds4 runtime that handles per-layer quant type dispatch. If you encounter "expected IQ2_XXS expert tensors" errors, use the uniform variant.
Domain Split (Calibration)
8,000 prompts · 5.0M tokens · 1.3B routed expert observations
| Domain | Share |
|---|---|
| Coding & development | 35-40% |
| Agentic tool-calling | 16% |
| Research & knowledge | 15-20% |
| Math & science | 10-15% |
| Design & planning | 5-10% |
| Trivia & general QA | 3-5% |
Calibration used the REAP activation_energy_sum2 score metric with 4,096 token context per prompt.
Top-to-bottom expert score gap in layer 3: 2,200x (strong pruning signal).
How to Run
Requires eouya2/ds4-for-reaped (ds4 engine with compact GGUF support):
git clone https://github.com/eouya2/ds4-for-reaped
cd ds4-for-reaped
make cuda-spark -j$(nproc) # DGX Spark / CUDA
# or: make # Metal / macOS
./ds4 --cuda -m DeepSeek-V4-Flash-REAP-K128.gguf --ctx 131072
API server mode:
./ds4-server --cuda -m DeepSeek-V4-Flash-REAP-K128.gguf \
--host 0.0.0.0 --port 17777 --ctx 131072
How It Was Built
- Donor GGUF: Downloaded antirez Layers37-42Q4KExperts variant (90.9 GiB) — Q4_K experts in deep layers, IQ2_XXS/Q2_K elsewhere
- Calibration: 8,000 prompts collected and run through ds4's imatrix collector on a DGX Spark (NVIDIA GB10) at 4,096 token context
- REAP scoring: Imatrix activation data converted to per-expert REAP scores using
activation_energy_sum2 - Pruning: 50% expert removal via
ds4_prune_gguf.pyfrom eouya2/reap-for-ds4. Layers 0-2 (hash-routed) preserved. Expert tensors copied byte-for-byte — no dequant/requant. - Output:
ds4-compact-v1GGUF
No fine-tuning. Purely structural expert removal. Weights are unmodified — a subset of the original experts.
Acknowledgments
- DeepSeek — DeepSeek V4 Flash base model
- antirez — ds4 engine and GGUF quants
- eouya2 — ds4-for-reaped and reap-for-ds4
- Cerebras Research — REAP (code)
- NVIDIA — DGX Spark hardware