---
language:
  - en
  - ko
  - zh
license: other
license_name: modified-mit
license_link: https://huggingface.co/moonshotai/Kimi-K2.7-Code/blob/main/LICENSE
tags:
  - gguf
  - kimi
  - moonshot
  - quantized
  - batiai
  - mixture-of-experts
  - coding
  - agentic
  - frontier
base_model: moonshotai/Kimi-K2.7-Code
pipeline_tag: text-generation
library_name: llama.cpp
---

# Kimi-K2.7-Code GGUF — Quantized by BatiAI

<p align="center">
  <a href="https://flow.bati.ai"><img src="https://img.shields.io/badge/BatiFlow-on--device%20AI-blue?style=for-the-badge&logo=apple" alt="BatiFlow"></a>
  <a href="https://huggingface.co/moonshotai/Kimi-K2.7-Code"><img src="https://img.shields.io/badge/source-Moonshot%20official-orange?style=for-the-badge" alt="moonshot"></a>
  <a href="#"><img src="https://img.shields.io/badge/1T--A32B-MoE-purple?style=for-the-badge" alt="MoE"></a>
</p>

> **The coding upgrade to Kimi K2.6 — +21.8% on Kimi Code Bench v2, running on a 512GB Mac Studio.**
> **IQ3_XXS / IQ4_XS** GGUF of **[moonshotai/Kimi-K2.7-Code](https://huggingface.co/moonshotai/Kimi-K2.7-Code)** (1T total / 32.6B active MoE, DeepSeek-V3-family architecture).
> Quantized **directly from official Moonshot weights** — code+multilingual imatrix, BatiAI-signed.

## 📦 Quantizations

| Quant | Size | Shards | Target |
|-------|------|--------|--------|
| **IQ3_XXS** | 394 GB (GiB: 367) | 10 | **M3 Ultra 512GB Mac Studio** |
| **IQ4_XS** | 546 GB (GiB: 509) | 13 | 512GB+ / multi-node / server |

Both built from official weights via a Q8_0 intermediate, quantized with a **code + EN + KO + ZH imatrix** (included: `Kimi-K2.7-Code-imatrix.dat`). Text-only (the vision tower of the K2.5-family checkpoint is not included; same as other K2 GGUFs).

**✅ Verified (this build, IQ3_XXS)** — captured greedy runs:
- Math: `127+58` → **185** (clean reasoning trace)
- Korean: 서울 소개 + 김치·비빔밥·불고기 각 한 문장 — fluent, zero token-mixing or loops
- Tool-call: `{"tool":"get_weather","args":{"city":"부산"}}` — exact JSON

## 🚀 Usage (llama.cpp — mainline, no fork needed)

```bash
hf download batiai/Kimi-K2.7-Code-GGUF "Kimi-K2.7-Code-IQ3_XXS-*.gguf" --local-dir ./k27

# llama.cpp auto-loads all shards from the first one
./llama-cli -m ./k27/Kimi-K2.7-Code-IQ3_XXS-00001-of-00010.gguf -ngl 99 -c 16384 \
  -p "Refactor this function and add tests."
```

Recommended sampling (Moonshot): `--temp 1.0 --top-p 0.95` (thinking mode). Architecture is `deepseek2` — supported by **mainline llama.cpp** out of the box. Ollama tags (`batiai/kimi-k2.7-code`) follow shortly.

## 📜 License

Modified MIT (Moonshot) — commercial use permitted; products exceeding 100M MAU / $20M monthly revenue must display "Kimi K2.7" attribution. Full text at the base model repo. Quantized weights redistributed under the same terms.

## ✨ What BatiAI did

- Direct from official Moonshot weights (never a re-quant of third-party GGUFs)
- Q8_0 intermediate + diverse imatrix (code/EN/KO/ZH) for balanced fidelity
- Verified: load ✅ · math ✅ · Korean ✅ · tool-call JSON ✅ — BatiAI metadata-signed

— **BatiAI** · on-device frontier AI · https://flow.bati.ai