DeepSeek-V4-Flash DSpark support GGUF (for the ds4 engine)

A standalone DSpark speculative-decoding draft ("support") model, packaged as a

single 5.6 GiB GGUF for the ds4 engine. It is the

drafter only — it does not generate on its own. Pair it with a DeepSeek-V4-Flash

main-model GGUF and ds4 proposes-and-verifies, committing only accepted prefixes.

This file is what ./download_model.sh dspark-support produces, but pre-built: you can

skip the ~167 GB DSpark source download and the local conversion.

What it is

DSpark is DeepSeek's official auxiliary draft module for DeepSeek-V4-Flash: three MTP

(multi-token-prediction) stages that read the main model's hidden states at layers 40–42

and propose a block of future tokens, with a Markov head for the block continuation. ds4

verifies every proposal against the main model, so **accepted tokens are exactly what the

main model would have produced** — speculative decoding is a speed optimization, not a

quality change.

Files

| File | Size | SHA-256 |

|---|---|---|

| DeepSeek-V4-Flash-DSpark-support.gguf | 5,989,114,272 B (5.6 GiB) | 8b3adf5942bec22ae2ea867cd7079cf13530ba83ffcffaf00f5de48664a1a34e |

• GGUF v3, 81 tensors, general.architecture = deepseek4-dspark.
• DSpark metadata: block_size = 5, markov_rank = 256, n_layers = 3,

target_layer_ids = [40, 41, 42], noise_token_id = 128799.

• Quantization (mixed, from the ds4 deepseek4-quantize DSpark builder): routed experts

ffn_down Q2_K / ffn_gate,ffn_up IQ2_XXS; attention, shared experts and main_proj

Q8_0; norms F32; hyper-connection tensors F16.

Usage

Requires a DeepSeek-V4-Flash main-model GGUF (e.g. ds4flash.gguf) and a recent ds4 build.

./ds4 -m ds4flash.gguf \
  --mtp DeepSeek-V4-Flash-DSpark-support.gguf \
  --dspark --temp 0

Notes:

• --dspark activates the runtime; --dspark-confidence 0.9 is the default acceptance

threshold (--dspark-confidence 0 forces fixed five-token blocks for diagnostics).

• DSpark requires greedy decoding (--temp 0); sampled and reasoning/think modes

disable it.

• It helps most on predictable continuations (e.g. code); low-yield prompts may see no

gain. Upstream reports draft acceptance ≈ 81–91% on structured/tool output and ≈ 60–68%

on general prose — those are DeepSeek/ds4 figures, not measured in this repository.

Provenance & method

1. Source: deepseek-ai/DeepSeek-V4-Flash-DSpark

(MIT) — the DeepSeek-V4-Flash checkpoint plus the DSpark drafter module. Only the

drafter tensors (the mtp.* stages + Markov head, checkpoint shards 46–48) were used.

2. Conversion: ds4's gguf-tools/deepseek4-quantize DSpark support builder, from

antirez/ds4 commit fc9efd1 ("Add DSpark speculative

decoding"). No engine code was modified to produce this file.

3. Verification: GGUF metadata read back and checked against the source config

(block_size, markov_rank, target_layer_ids, noise_token_id).

This is a straightforward format conversion of MIT-licensed weights. It has been built and

metadata-verified; end-to-end acceptance/throughput on your specific hardware should be

confirmed with ds4's make dspark-acceptance fixture.

License & attribution

MIT, inherited from the base model. This is a derivative of DeepSeek's checkpoint and adds

no additional restrictions. Not affiliated with or endorsed by DeepSeek or the ds4 author;

all credit for the DSpark method and weights goes to DeepSeek, and to antirez for the ds4

engine and its DSpark GGUF builder.