Uncensored DeepSeek-V4-Flash-0731 — safety refusals removed while fully preserving knowledge, reasoning (controllable effort), tool-calling, and DSpark speculative decoding. Native FP8 + FP4 (NVFP4-class) checkpoint (~167 GB), built to serve tensor-parallel across two NVIDIA GB10 / DGX Spark nodes.

Research artifact with reduced safety guardrails. Use responsibly and lawfully.

Model

Base: deepseek-ai/DeepSeek-V4-Flash-0731 (43-layer MoE, 256 routed + 1 shared expert, MLA, sliding-window attention, ~1M context via YaRN). Native quant: FP8 (attention + shared experts)
- FP4 (routed experts), DSpark MTP head — all preserved. Serves unchanged in the DeepSeek-V4 vLLM path.

Benchmarks (vs. base, same checkpoint, on 2×GB10 TP2)

Metric	Base	This model
MMLU (logit, 285Q)	84.2%	86.7% (+2.5pp)
HumanEval pass@2 (70)	—	94.3%
HarmBench compliance (240)	refuses	≥96%, 0 hard refusals
Decode throughput (single-stream)	—	45–49 tok/s
Prefix-cache hit (12-way concurrent)	—	64% (3.7× faster on hit)
Multiturn correctness (24 concurrent)	—	24/24
GPU KV cache	—	1.27M tokens (fp8)

HarmBench compliance by category

240 behaviors, answer-channel (no-refusal) compliance, greedy:

Category	Comply	Rate
chemical biological	42/42	100%
cybercrime intrusion	52/52	100%
harassment bullying	21/21	100%
harmful	18/18	100%
illegal	53/53	100%
misinformation disinformation	54/54	100%

Knowledge and reasoning are fully retained — MMLU is within noise (slightly higher). Compliance is measured on the answer channel.

Recommended sampling

Official DeepSeek-V4-Flash defaults, stamped as this repo's generation_config.json:

Param	Value
`temperature`	1.0
`top_p`	0.95

{"do_sample": true, "temperature": 1.0, "top_p": 0.95}

These are applied automatically when a request omits sampling params. Override per-request as needed.

Serving (2× DGX Spark GB10, tensor-parallel-2)

Requires the DeepSeek-V4 vLLM path for sm_121 (e.g. the jasl/vllm GB10 build). Head + worker over a RoCE point-to-point link, TP=2 + expert-parallel, fp8 KV cache, DSpark spec-decode, 384K ctx. See the included serve/ scripts. Key flags:

vllm serve <this-model> --tensor-parallel-size 2 --enable-expert-parallel \
  --distributed-executor-backend mp --nnodes 2 --node-rank {0,1} --master-addr <head> \
  --kv-cache-dtype fp8 --block-size 256 --enable-prefix-caching \
  --max-model-len 393216 --gpu-memory-utilization 0.80 \
  --speculative-config '{"method":"dspark","num_speculative_tokens":5}' \
  --reasoning-parser deepseek_v4 --tool-call-parser deepseek_v4 --enable-auto-tool-choice

Reasoning effort

Controllable thinking (low / high / max) via the reasoning system preface; the reasoning trace is returned as reasoning_content, the answer as content. Note: thinking mode reasons at length — allow ≥2500 output tokens or the answer may be truncated.

Tools

DeepSeek-V4 tool-call parser; pass OpenAI-style tools to the chat endpoint.

License

MIT (inherits the upstream DeepSeek-V4-Flash license terms). Review before use.

Contact

eric@dealign.ai

dealignai/DeepSeek-V4-Flash-0731-CRACK-NVFP4

DeepSeek-V4-Flash-0731-CRACK-NVFP4