--- license: other base_model: moonshotai/Kimi-K2.7-Code tags: - gguf - llama.cpp - kimi - kimi-k2.7-code - bf16 - quantized - tq1_0 - tq2_0 - q4_k_m - q8_0 library_name: gguf --- # Kimi K2.7 Code GGUF GGUF conversions of [`moonshotai/Kimi-K2.7-Code`](https://huggingface.co/moonshotai/Kimi-K2.7-Code) for llama.cpp-compatible runtimes. ## Available files | Variant | Files | Approx local size | Status | Notes | |---|---:|---:|---|---| | Q8_0 | `Q8_0/kimi-k2.7-code-Q8_0-00001-of-00061.gguf` ... `00061-of-00061.gguf` | ~1017 GiB | Uploaded | 8-bit GGUF, split into 61 shards. | | Q4_K_M | `Q4_K_M/kimi-k2.7-code-Q4_K_M-00001-of-00061.gguf` ... `00061-of-00061.gguf` | ~578 GiB | Uploaded | Standard high-quality 4-bit GGUF, split because the single file exceeds Hugging Face’s 500GB per-file limit. | | TQ2_0 | `TQ2_0/kimi-k2.7-code-TQ2_0.gguf` | ~249 GiB | Uploaded | 2-bit-class ternary quantization, single file. | | TQ1_0 | `TQ1_0/kimi-k2.7-code-TQ1_0.gguf` | ~204 GiB | Uploaded | 1-bit-class ternary quantization, single file. | `Q6_K` is not currently uploaded in this repository. ## Loading split GGUF files For split GGUF variants, download all shards for the variant into the same directory and point llama.cpp at shard `00001`. llama.cpp will discover the remaining shards automatically. Examples: ```bash # BF16 llama-cli -m BF16/kimi-k2.7-code-BF16-00001-of-00061.gguf -p "Write a Python function for quicksort." # Q8_0 llama-cli -m Q8_0/kimi-k2.7-code-Q8_0-00001-of-00061.gguf -p "Write a Rust HTTP server." # Q4_K_M llama-cli -m Q4_K_M/kimi-k2.7-code-Q4_K_M-00001-of-00061.gguf -p "Explain async/await." ``` Single-file variants can be loaded directly: ```bash llama-cli -m TQ2_0/kimi-k2.7-code-TQ2_0.gguf -p "Hello" llama-cli -m TQ1_0/kimi-k2.7-code-TQ1_0.gguf -p "Hello" ``` ## Quantization notes - `BF16` was converted from the original SafeTensors using llama.cpp `convert_hf_to_gguf.py` with BF16 output. - `Q8_0` and `Q4_K_M` were quantized from the BF16 GGUF source and uploaded as split GGUF shards. - `TQ1_0` and `TQ2_0` are llama.cpp ternary low-bit formats. - `IQ1_S` was not produced because llama.cpp requires an importance matrix for that quantization. - Very large variants are split to stay under Hugging Face’s individual file-size limit. ## License See [`LICENSE`](./LICENSE). This model uses Moonshot AI’s Modified MIT License for Kimi K2.7 Code. ## Attribution Base model by Moonshot AI: [`moonshotai/Kimi-K2.7-Code`](https://huggingface.co/moonshotai/Kimi-K2.7-Code).
