[!TIP] Support this work → · X · GitHub · REAP paper · Cerebras REAP
DeepSeek-V4-Flash-162B-GGUF
GGUF quantization of 0xSero/DeepSeek-V4-Flash-162B.
At a glance
| Base model | 0xSero/DeepSeek-V4-Flash-162B |
| Format | GGUF |
| Total params | 162B |
| Active / token | — |
| Experts / layer | — |
| Layers | — |
| Hidden size | — |
| Context | — |
| On-disk size | 149 GB |
Which variant should I pick?
| Variant | Format | Link |
|---|---|---|
DeepSeek-V4-Flash-162B | BF16 | link |
DeepSeek-V4-Flash-162B-GGUF (this) | GGUF | link |
DeepSeek-V4-Flash-180B | BF16 | link |
DeepSeek-V4-Flash-180B-GGUF | GGUF | link |
DeepSeek-V4-Flash-213B | BF16 | link |
This repository contains DS4/DwarfStar GGUF conversions of DeepSeek-V4-Flash-Spark-Mini.
The GGUFs point back to the original Spark Hugging Face model:
- Original Spark model: https://huggingface.co/0xSero/DeepSeek-V4-Flash-162B
- Conversion source checkpoint: https://huggingface.co/0xSero/DeepSeek-V4-Flash-162B-codex-K144-REAP
- Runtime/converter repo: https://github.com/antirez/ds4
- Spark deployment repo: https://github.com/0xSero/deepseek-spark
Files
| File | Size | SHA256 |
|---|---|---|
DeepSeek-V4-Flash-Spark-Mini-Q2-REAP-ds4.gguf | 48.98 GiB | e917278028d7a9e25dfc9d04bf5848375dad7573c5aeab1720d6a83714352406 |
Quantization
Q2-REAP-ds4: compact DS4 profile usingIQ2_XXSrouted gate/up experts,Q2_Krouted down experts, andQ8_0shared/output/attention projections.
These are DS4/DwarfStar-specific GGUF files for DeepSeek-V4 Flash REAP checkpoints. They are not generic llama.cpp files unless your runtime supports the same DeepSeek-V4 Flash tensor layout and DS4 metadata.
Validation
Validation summaries are uploaded in this repo under:
validation/20260528T160633Z/SUMMARY.mdvalidation/20260528T160633Z/summary.json
The Mini Q2 GGUF completed the DS4 context sweep through 200000 context on one DGX Spark:
| Context | Prefill tok/s | Decode tok/s | KV bytes |
|---|---|---|---|
| 2,048 | 348.19 | 12.75 | 52,184,460 |
| 4,096 | 358.51 | 13.50 | 80,373,132 |
| 8,192 | 352.29 | 13.32 | 136,750,476 |
| 16,384 | 348.25 | 13.24 | 249,505,164 |
| 32,768 | 322.07 | 12.40 | 475,014,540 |
| 65,536 | 287.26 | 11.49 | 926,033,292 |
| 131,072 | 241.57 | 9.81 | 1,828,070,796 |
| 200,000 | 194.24 | 9.17 | 2,776,775,308 |
API probes completed through at least the 131072 window before spark-2822 became unreachable during the tail of the 200000 validation step:
| Context | Prompt tokens | TTFT seconds | Prefill tok/s | Decode tok/s | Marker visible |
|---|---|---|---|---|---|
| 65,536 | 59,867 | 176.54 | 339.12 | 13.01 | true |
| 131,072 | 119,696 | 390.59 | 306.45 | 11.70 | true |
This repo publishes the validated Q2 long-context profile only.
License & citation
License inherited from the base model.
@misc{lasby2025reap,
title = {REAP the Experts: Why Pruning Prevails for One-Shot MoE Compression},
author = {Mike Lasby and Ivan Lazarevich and Nish Sinnadurai and Sean Lie and Yani Ioannou and Vithursan Thangarasa},
year = {2025}, eprint = {2510.13999}, archivePrefix = {arXiv}
}
Sponsors
Made possible by NVIDIA · TNG Technology · Lambda · Prime Intellect · Hot Aisle.