DeepSeek-V4-Flash-DSpark

DeepSeek-V4-Flash-DSpark is the 284B-parameter (13B-activated) Mixture-of-Experts member of the DeepSeek-V4 family, with a 1-million-token context window and FP8 mixed-precision weights. The -DSpark variant attaches a native Multi-Token-Prediction (MTP) speculative-decoding draft head (DeepSpec).

Evaluation

Benchmark	Full N	Score
MMLU-Pro	12032	0.6750
GSM8K	1319	0.9257
HumanEval (pass@1)	164	0.8354
MBPP (pass@1)	500	0.5160

Multi-turn & higher-context evaluation

Benchmark	Score
Multi-turn (20 curated 3-turn convos / 60 turns, Gemini judge 1-10)	9.98 / 10
Needle-in-haystack @ 2k / 4k / 8k / 16k / 32k tokens	100% / 100% / 100% / 100% / 100%

SWE-bench Lite (oracle-file-context, single-shot, n=30)

Metric	Value
Resolved	4 / 30 (13.3%)
Completed	19 / 30
Patch-apply errors	11 / 30

Speculative decoding throughput

Single-stream: ~238 tok/s at temperature=0 (TP=2, DSpark native 5-token speculation, lucifer-cutlass backend).

Safety compliance by category (1000-prompt eval set)

Category	Compliance	Retention
PII (doxing private individuals)	42 / 78	53.8%
Self-Harm (suicide methods)	59 / 70	84.3%
Radicalization	62 / 67	92.5%
Cybercrime	63 / 67	94.0%
Hate Speech	74 / 78	94.9%
Illegal Drugs	66 / 69	95.7%
Weapons	70 / 72	97.2%
Political Sensitivity	67 / 69	97.1%
Fraud	76 / 78	97.4%
Harassment	59 / 61	96.7%
Violence	75 / 76	98.7%
CBRNE	69 / 70	98.6%
Financial Crimes	70 / 71	98.6%
Sabotage	74 / 74	100.0%
Overall (unweighted)	926 / 1000	92.6%

Safety guardrails are highest in PII doxing (46.2% retention) and self-harm (84.3%), where the model defaults to empathetic hotline-style responses and refuses to surface private contact data. In all 12 remaining categories compliance exceeds 92%.

Files

All 48 safetensors shards are included, plus model.safetensors.index.json, config.json, generation_config.json, tokenizer.json, tokenizer_config.json, LICENSE, and the encoding/ and inference/ folders. It loads directly with vLLM / the DeepSeek-V4 inference path.

Usage

This checkpoint is a drop-in replacement for the original weights — it has the same architecture, format, chat-template/encoding, and inference path as deepseek-ai/DeepSeek-V4-Flash-DSpark. Load and serve it with vLLM or the DeepSeek-V4 inference stack.