RadixArk/Kimi-K3-DSpark

RadixArk/Kimi-K3-DSpark is listed on Hugging Face as a text-generation model. This page records the public repository snapshot 3c5bac301d9cf392706189d82ed947feca6c2f0f; verify the listed revision, licence, and runtime support before using it.

Updated

Computraq is preparing for launch. These guides explain readiness and the proposed framework; they do not establish a live service, integration or partnership.

Identity and access

The publisher record identifies this repository as RadixArk/Kimi-K3-DSpark. Its current public snapshot is 3c5bac301d9cf392706189d82ed947feca6c2f0f. Access is public repository; review the applicable licence and publisher terms; the recorded licence label is not stated in the public record. A public repository or gated download does not establish permission for every use.

Configuration facts

The pinned configuration reports model type qwen3 and architecture DSparkDraftModel. The model hub tags it for text-generation; these labels describe repository metadata and do not establish benchmark performance.

  • head_dim: 64
  • hidden_size: 7168
  • intermediate_size: 14336
  • max_position_embeddings: 1048576
  • num_attention_heads: 64
  • num_hidden_layers: 5
  • num_key_value_heads: 16
  • rms_norm_eps: 1e-05
  • rope_parameters.rope_theta: 10000.0
  • vocab_size: 163840

Model-specific qualification

The pinned architecture and speculative-decoding tag identify this as a draft-model component. Verify its target-model pairing, tokenizer compatibility, decoding protocol, and serving-runtime support from the publisher’s documentation before use; do not assume this checkpoint works as a standalone chat model.

Weight-memory planning

This guide does not retain an independently verified logical parameter count, so it omits a numeric weight-memory estimate. If a count is verified for this exact repository revision, use parameter count × bytes per parameter for a weights-only floor; then account separately for KV cache, context, activations, runtime, and concurrency.

Revision and reproducibility

The metadata and configuration are pinned to revision 3c5bac301d9cf392706189d82ed947feca6c2f0f. Pin this commit when reproducing an evaluation. Repository branches can change, and this source check does not validate the model weights themselves.

Published record

  • Task: text-generation
  • Model type: qwen3
  • Architecture: DSparkDraftModel
  • Licence label: not stated in the public record
  • Access: public repository; review the applicable licence and publisher terms
  • Revision: 3c5bac301d9cf392706189d82ed947feca6c2f0f

Sources and verification