Modelpublic

0xSero/DeepSeek-V4-Flash-180B-GGUF

sourceHugging Facemitupdated 4mo agoView on Hugging Face

14likes140downloads

Model Card

[!TIP] [Support this work →](https://donate.sybilsolutions.ai) · X · GitHub · REAP paper · Cerebras REAP

DeepSeek-V4-Flash-180B-GGUF

GGUF quantization of 0xSero/DeepSeek-V4-Flash-180B.

At a glance


Base model	0xSero/DeepSeek-V4-Flash-180B
Format	GGUF
Total params	180B
Active / token	—
Experts / layer	—
Layers	—
Hidden size	—
Context	—
On-disk size	164 GB

Which variant should I pick?

Variant	Format	Link
`DeepSeek-V4-Flash-162B`	BF16	link
`DeepSeek-V4-Flash-162B-GGUF`	GGUF	link
`DeepSeek-V4-Flash-180B`	BF16	link
`DeepSeek-V4-Flash-180B-GGUF` (this)	GGUF	link
`DeepSeek-V4-Flash-213B`	BF16	link

This repository contains DS4/DwarfStar GGUF conversions of DeepSeek-V4-Flash-Spark.

The GGUFs point back to the original Spark Hugging Face model:

—Original Spark model: https://huggingface.co/0xSero/DeepSeek-V4-Flash-180B
—Conversion source checkpoint: https://huggingface.co/0xSero/DeepSeek-V4-Flash-180B-codex-K160-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-Q2-REAP-ds4.gguf`	53.52 GiB	`dae2ed196e8ad87d6667d3fa04f65d78302ea4f148ed0ee0f3ff0b829d1f9c5d`

Quantization

—Q2-REAP-ds4: compact DS4 profile using IQ2_XXS routed gate/up experts, Q2_K routed down experts, and Q8_0 shared/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.md
—validation/20260528T160633Z/summary.json

The Spark 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	360.26	13.63	52,184,460
4,096	357.05	13.74	80,373,132
8,192	360.20	13.56	136,750,476
16,384	348.30	13.31	249,505,164
32,768	333.74	12.59	475,014,540
65,536	306.45	11.79	926,033,292
131,072	267.63	10.29	1,828,070,796
200,000	214.11	9.25	2,776,775,308

The corrected 200K API probe used 182,633 prompt tokens and returned the visible marker SPARK-CTX-200000-OMEGA:

Prompt tokens	TTFT seconds	Prefill tok/s	Decode tok/s	Passed core
182,633	668.53	273.19	10.61	true

Terminal-Bench 2.0 evidence is included in the validation summary: one real gpt2-codegolf trial completed without harness errors after enabling amd64 binfmt on the ARM64 Spark host.

This repo publishes the validated Q2 long-context profile only.

License & citation

License inherited from the base model.

bibtex

@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}
}