---
language:
  - en
license: other
license_name: modified-mit
license_link: https://huggingface.co/moonshotai/Kimi-K2.7-Code/blob/main/LICENSE
base_model: moonshotai/Kimi-K2.7-Code
tags:
  - gguf
  - kimi
  - kimi-k2
  - code
  - agent
  - moe
  - multimodal
  - vision
  - llama-cpp
  - image-text-to-text
library_name: gguf
pipeline_tag: image-text-to-text
inference: true
model_name: Kimi K2.7-Code (GGUF — server-class quants)
---

# Kimi K2.7-Code — GGUF (coding agent MoE)

**Community GGUF mirror** of [moonshotai/Kimi-K2.7-Code](https://huggingface.co/moonshotai/Kimi-K2.7-Code) for **llama.cpp**-compatible runtimes on **server-grade hardware**.

Released **June 12, 2026** by Moonshot AI. Coding-focused agent built on Kimi K2.6 with +21.8% on Kimi Code Bench v2.

| | |
|---|---|
| **Architecture** | 1T MoE (32B active), DeepSeek2 / MLA |
| **Context** | **256K** tokens (262144 in GGUF) |
| **Modalities** | Text, **image**, **video** (API-first; vision via mmproj in GGUF) |
| **License** | Modified MIT |
| **Thinking** | Forced `preserve_thinking` — reasoning retained across turns |

## Important: server-class model

This is **not** a consumer-laptop model. Even the smallest GGUF quants are **hundreds of GB**. Plan for:

- Multi-GPU or high-RAM server (512 GB+ system RAM typical for Q4-class quants)
- Fast NVMe scratch space
- Latest **llama.cpp** with DeepSeek2 / Kimi K2.5+ support

See [docs/kimi-k27-code-analysis.md](https://github.com/Edmon02/audio_set/blob/main/docs/kimi-k27-code-analysis.md) for full analysis.

## Why this repo exists

- **One download hub** for unsloth UD quants (Q2–Q8, IQ variants) + mmproj.
- **Hub-side sync** from [unsloth/Kimi-K2.7-Code-GGUF](https://huggingface.co/unsloth/Kimi-K2.7-Code-GGUF) — no re-upload from your laptop.
- Maintainer script: `scripts/sync_kimi_k27_code_gguf_quants.py`

## Available files

See [`gguf-manifest.json`](gguf-manifest.json) for the live file list.

### Essential tier (recommended start)

| Path | Use |
|------|-----|
| `UD-Q4_K_XL/` (14 shards) | **Recommended** — maps to Kimi native int4 quality |
| `mmproj-F16.gguf` | Vision encoder weights for llama.cpp multimodal |
| `config.json` | Model metadata |

### Full tier

All unsloth UD quants (`UD-IQ1_M`, `UD-IQ3_XXS`, `UD-IQ4_XS`, `UD-Q2_K_XL`, `UD-Q3_K_XL`, `UD-Q4_K_XL`, `UD-Q8_K_XL`) + mmproj BF16/F16/F32 — run `make sync-kimi-k27-gguf-full`.

## Download

```bash
pip install -U huggingface_hub

# Essential: Q4 XL + vision mmproj (hundreds of GB)
huggingface-cli download Edmon02/Kimi-K2.7-Code-GGUF \
  config.json mmproj-F16.gguf \
  --include "UD-Q4_K_XL/*" \
  --local-dir ./models/kimi-k27-code
```

## Quick start (llama.cpp)

Requires a recent llama.cpp build with Kimi K2.5 / DeepSeek2 MoE support.

```bash
# Text + tools (thinking mode — match Moonshot API defaults)
llama-server -m ./models/kimi-k27-code/UD-Q4_K_XL \
  --mmproj ./models/kimi-k27-code/mmproj-F16.gguf \
  --ctx-size 32768 \
  --temp 1.0 --top-p 0.95
```

Moonshot recommends **temperature=1.0**, **top_p=0.95**, and **thinking enabled**. Instant mode is not supported.

## Benchmark highlights (Moonshot-reported)

| Benchmark | K2.6 | **K2.7-Code** | Δ vs K2.6 |
|-----------|------|---------------|-----------|
| Kimi Code Bench v2 | 50.9 | **62.0** | +21.8% |
| Program Bench | 48.3 | **53.6** | +11.0% |
| MLS Bench Lite | 26.7 | **35.1** | +31.5% |
| MCP Atlas | 69.4 | **76.0** | +9.5% |
| MCP Mark Verified | 72.8 | **81.1** | +11.4% |

## Deployment alternatives

| Path | When |
|------|------|
| **Kimi API** (`kimi-k2.7-code`) | Production agents, Kimi Code CLI |
| **vLLM / SGLang / KTransformers** | Self-host from safetensors |
| **GGUF + llama.cpp** | Offline / custom infra with enough RAM |

API pricing (Moonshot): ~$0.95 / $4.00 per 1M tokens in/out.

## Provenance

| Item | Source |
|------|--------|
| Base model | `moonshotai/Kimi-K2.7-Code` |
| GGUF quants | Mirrored from `unsloth/Kimi-K2.7-Code-GGUF` |
| Maintainer | [Edmon02/audio_set](https://github.com/Edmon02/audio_set) |

## Limitations

- Sharded GGUF folders — download entire quant prefix, not individual shards only.
- Video input in GGUF may lag official API support.
- Vendor-run benchmarks; validate on your coding/agent workloads.
- GGUF community quants — compare against native int4 safetensors when possible.

## Citation

```bibtex
@misc{kimi_k27_code_2026,
  title={Kimi K2.7-Code},
  author={Moonshot AI},
  year={2026},
  url={https://huggingface.co/moonshotai/Kimi-K2.7-Code}
}
```