Kimi-K3-8L-dummy

An 8-layer, config-only variant of moonshotai/Kimi-K3 for testing and debugging inference engines. It contains no weights.

Usage

Load it with random weights, e.g. in vLLM:

vllm serve riverclouds/Kimi-K3-8L-dummy \
  --trust-remote-code \
  --load-format dummy \
  --tensor-parallel-size 4

Outputs are meaningless. It is meant for exercising the real Kimi-K3 architecture (KDA linear attention + MLA, MoE, KV-cache layout, scheduling, prefix caching) at a fraction of the size.

Changes from the official config

Only three fields in text_config differ from moonshotai/Kimi-K3's config.json:

Field	Kimi-K3	This repo
num_hidden_layers	93	8
linear_attn_config.kda_layers	[1, 2, 3, 5, ..., 91]	[1, 2, 3, 5, 6, 7]
linear_attn_config.full_attn_layers	[4, 8, ..., 92, 93]	[4, 8]

This keeps the 3:1 KDA-to-MLA ratio, with a full-attention last layer as in the original. All other fields (including max_position_embeddings, auto_map, and vision_config) are unchanged. Tokenizer, processor, generation config and remote code files are copied verbatim from the official repo.

License

Distributed under the original Kimi K3 License from Moonshot AI.