Modelpublic

freakyskittle/kimi-k2.75-code-GGUF

sourceHugging Faceotherupdated 3mo agoView on Hugging Face

1likes148downloads

Model Card

Kimi K2.75 Code — GGUF

GGUF conversions and quantizations of the experimental merged checkpoint `freakyskittle/kimi-k2.75-code` (a shard-wise SLERP merge of moonshotai/Kimi-K2.7-Code and moonshotai/Kimi-K2.6, DeepSeek-V3 style MoE).

See the base repository for the full merge recipe, pruning details, and license.

Parameter count: ~720 B. The bf16 and Q80 builds contain 1,096 tensors summing to **719.8 B** parameters. Hugging Face's auto-detected badge reads ~61 B because it parses the `deep55/` file first, and that build stores weights in a packed-int4 layout (int32 `weightpacked tensors holding 8 weights each, plus weightscale`/`weightshape) whose logical shapes can't be summed. The true model is ~720 B params (further reduced in the deep-pruned deep55/` variant).

Files

Each variant lives in its own folder. Files over Hugging Face's 500 GB per-file limit are split into GGUF shards (-0000N-of-0000M.gguf); point your loader at the first shard and the rest are picked up automatically.

Folder	Variant	Approx. size	Notes
`deep55/`	Deep-pruned, full precision	~203 GB	Prune ratio `0.55`, `deepseek_v3` arch override
`pruned-compact-oxidize-q4/`	Q4KM (oxidize)	~406 GB	Quantized from the compacted pruned checkpoint
`llamacpp-q4-partial/`	Q4KM (llama.cpp, partial)	~435 GB	Partial llama.cpp quantization
`unpruned-q8/`	Q8_0 (unpruned)	~765 GB	Split into 2 shards
`pruned-bf16/`	BF16 (pruned)	~1.44 TB	Split into 4 shards

Provenance

—Base merge + pruning: see `freakyskittle/kimi-k2.75-code`.
—GGUF conversion via oxidize-convert with --arch deepseek_v3.
—Sharding via llama-gguf-split (--split-max-size 450G).

Status

Experimental research artifact — not fully evaluated. Validate quality before any production use. Use at your own risk.

License

Follows the Modified MIT License from Moonshot AI (see the base repository). Commercial attribution requirement applies: products/services exceeding 100M monthly active users or US$20M monthly revenue must prominently display Kimi K2.7 Code in the UI.