---

base_model: RadixArk/Kimi-K3-DSpark

library_name: gguf

pipeline_tag: text-generation

tags:

• gguf
• q8_0
• speculative-decoding
• dspark
• dflash
• kimi-k3
• lucebox

inference: false

---

Kimi K3 DSpark Q8_0 GGUF

Reproducible Q8_0 conversion of

RadixArk/Kimi-K3-DSpark,

a DSpark speculative drafter trained for

moonshotai/Kimi-K3.

The artifact is intended for Lucebox's generic DFlash/DSpark runtime contract.

It contains the drafter only; Kimi K3 target weights, embeddings, and

unembedding weights are not included. Kimi-specific runtime integration is

still required before this is a drop-in speculative-decoding model in a Lucebox

release.

Conversion result

| Property | Value |

|---|---:|

| Source format | BF16 Safetensors |

| Source size | 4,498,585,858 bytes |

| Output format | GGUF MOSTLY_Q8_0 |

| Output size | 2,390,153,888 bytes |

| Output/source | 53.13% |

| Parameters | 2,249,289,601 |

| Q8_0 tensors | 38 |

| F32 tensors | 24 |

| Sampled relative RMSE, RMS | 0.00544127 |

| Maximum per-tensor sampled relative RMSE | 0.00557332 |

All large two-dimensional matrices, including the DFlash fusion projection and

both low-rank Markov matrices, are Q8_0. Norms and the small confidence head are

kept in F32. This cuts drafter storage by 46.9% while retaining high precision

for the small, sensitive tensors.

The RMSE figures validate the weight conversion, not end-to-end acceptance.

Acceptance length and speed must be measured after the Kimi K3 verifier,

hidden-state capture, ReplaySSM, and adaptive DSpark scheduling are integrated.

Integrity and provenance

• Source repository: RadixArk/Kimi-K3-DSpark
• Pinned source revision: 56ce616ad7486f0e96cbb51ef23ed5a1bce1d92d
• Source model.safetensors SHA256:

ecd746459b4a603ce0d2c64f73935efead29bd651b14439b99e57ee8b41b77ca

• Output Kimi-K3-DSpark-Q8_0.gguf SHA256:

848d12be5283a4717c08b1b1263420980dfc61b5a59067e343296fd7f4998435

conversion-report.json contains per-tensor sampled errors, tensor counts,

sizes, source revision, and both hashes. The same provenance and model geometry

are embedded in the GGUF metadata.

Architecture metadata

• Generic GGUF architecture: dflash-draft
• Draft blocks: 5
• Target blocks: 93
• Captured target layers: [7, 23, 51, 67, 83]
• DSpark block size: 7
• Hidden size: 7168
• Attention heads: 64 query / 16 KV, head dimension 64
• Markov rank: 256
• Context metadata: 1,048,576 with YaRN-16 from 65,536

The target block count and captured-feature count are stored separately. This

is important for Kimi K3: its target has 93 blocks, while fc.weight consumes

five captured hidden states.

Download

hf download Lucebox/Kimi-K3-DSpark-Q8_0-GGUF \
  Kimi-K3-DSpark-Q8_0.gguf conversion-report.json

Upstream attribution

The drafter was trained and published by RadixArk using SpecForge. See the

upstream model card for its

training procedure, evaluation results, and SGLang usage. No license field was

declared in that upstream model card at conversion time; downstream users

should verify the applicable upstream terms.