license: mit
base_model:
- apetersson/DeepSeek-V4-Flash-0731-Abliterated-FP8
pipeline_tag: text-generation
tags: - deepseek-v4
- nvfp4
- fp8
- modelopt
- quantized
- dgx-spark
DeepSeek V4 Flash 0731 Abliterated NVFP4 — archived conversion
This repo is now an informational page. The checkpoint weights that used
to live here have been removed (2026-08-16): the conversion is superseded
and there is nothing to download.
In plain English: this was a repackaging of DeepSeek V4 Flash (0731,
abliterated) with the routed-expert weights in NVIDIA NVFP4 format. Since
then, faster and better-supported paths won out — so instead of these
weights, use:
- NVIDIA DGX Spark (GB10):
drowzeys/keys-DeepSeekV4-Flash-GA-0731-Dspark-Abliterated-32-32— the DSpark-native MXFP4-path checkpoint. Fastest on Sparks in our tests. - The full 4× Spark TP=4 serving recipe, benches, and fabric runbook:
github.com/neko-legends/spark-bench — reproduce everything below. - Server-class Blackwell / RTX PRO stacks: the NVFP4 conversion notes
below remain for reference, but the weights are gone; run your own cast
with NVIDIA TensorRT Model Optimizer.
What four DGX Sparks do with the 32-32 checkpoint today
Single-stream decode (client wall, 2048-token completions, formal protocol):
The full ledger — decode, prefill, and concurrency:
| metric | TP2 baseline (07-31) | TP4 broken, no-spec (08-15) | TP4 record (08-14) | TP4 now (08-16) |
|---|---|---|---|---|
| C1 decode (tok/s) | 67.7 | 33.5 | 103.4 | 136.25 median · 145.5 peak |
| Prefill cold (tok/s) | 1576 | ~950 | ~940 | 2102 @32k · 2156 @8k |
| C4 aggregate (tok/s) | 93.15 | 92.43 | — | 182.2 |
Decode at prompt depth — speculative-decoding acceptance, not depth, is the
variable:
Concurrency-4 aggregate:
Everything needed to reproduce — image, flags, fabric wiring, boot gates,
bench scripts, dated results: github.com/neko-legends/spark-bench.
About the removed conversion (reference)
- Source:
apetersson/DeepSeek-V4-Flash-0731-Abliterated-FP8 — routed experts were packed MXFP4 E2M1 with per-32 UE8M0 scales; we used NVIDIA ModelOpt's lossless MXFP4→NVFP4 cast (no dequantize/requantize) with activation scales calibrated on 2× GB10 DGX Spark (128 public prompts, batch 4, seq 512). - Mixed precision by design: routed experts NVFP4 E2M1; attention/dense FP8 E4M3; shared-expert and MTP tensors retained.
- Export audit at release:
8,657,043,456 / 8,657,043,456blocks cast, 33,024 tensors across 43 layers, 48/48 shards independently audited, zero structural errors. - Abliteration retention was verified at release against the upstream ablation manifest (all 36 named tensors bit-identical) plus a small behavioral suite (12/12 substantive, zero refusals).
- The runtime bridge for the SM121 B12X path remains in
runtime/, the reference inference code ininference/, and the DSV4 tokenizer encoding inencoding/.
Attribution
- Base model: DeepSeek-AI DeepSeek V4 Flash 0731
- Abliterated source: apetersson/DeepSeek-V4-Flash-0731-Abliterated-FP8
- Quantization/export tooling: NVIDIA TensorRT Model Optimizer
- Conversion + serving recipe:
neko-legends(spark-bench)
Review the upstream model card and license before deploying any descendant
checkpoint.