What is Kimi-K2.7-Code?
Kimi-K2.7-Code is a coding-focused agentic model from Moonshot AI, built on top of Kimi K2.6. It is a 1T-parameter Mixture-of-Experts model that activates 32B parameters per token, tuned for long-horizon software engineering: finishing multi-step coding tasks end to end instead of answering one prompt at a time. Moonshot also cut thinking-token usage by roughly 30% compared with K2.6, so the same job burns fewer tokens. The weights landed on Hugging Face on June 11, 2026 under a Modified MIT license.
| Specification | Kimi-K2.7-Code |
|---|---|
| Total parameters | 1T |
| Activated parameters | 32B per token |
| Architecture | Mixture-of-Experts: 384 experts, 8 selected per token plus 1 shared, MLA attention |
| Context window | 256K tokens |
| Modalities | Text and image input, video input experimental |
| Vision encoder | MoonViT, 400M parameters |
| Reasoning | Thinking forced on, reasoning kept across turns |
| Native quantization | INT4 |
| Release date | June 11, 2026 |
| License | Modified MIT |
Per token the router picks 8 of the 384 experts plus 1 shared expert, so a forward pass does the work of a 32B model while the full 1T parameters sit in memory. Two design choices matter for agent work. Thinking is forced on and the model keeps its reasoning content across turns, a mode Moonshot calls preserve_thinking and says raises performance in coding-agent scenarios. And the weights ship in native INT4, the same quantization method as Kimi-K2-Thinking. The model carries a 400M-parameter MoonViT vision encoder, so it takes image input alongside text; video input exists but only as an experimental feature of Moonshot's official API.
Kimi-K2.7-Code benchmarks
Moonshot's model-card numbers compare K2.7 Code with its predecessor Kimi K2.6, GPT-5.5 running in Codex, and Claude Opus 4.8 running in Claude Code:
| Benchmark | Kimi-K2.7-Code | Kimi K2.6 | GPT-5.5 | Claude Opus 4.8 |
|---|---|---|---|---|
Kimi Code Bench v2 Realistic coding | 62.0 | 50.9 | 69.0 | 67.4 |
Program Bench Program recreation | 53.6 | 48.3 | 69.1 | 63.8 |
MLS Bench Lite ML research | 35.1 | 26.7 | 35.5 | 42.8 |
Kimi Claw 24/7 Bench Long-horizon agents | 46.9 | 42.9 | 52.8 | 50.4 |
MCP Atlas MCP tools | 76.0 | 69.4 | 79.4 | 81.3 |
MCP Mark Verified MCP servers | 81.1 | 72.8 | 92.9 | 76.4 |
K2.7 Code improves on K2.6 in all six rows while spending about 30% fewer thinking tokens, and it beats Claude Opus 4.8 on MCP Mark Verified. It takes no top scores though: GPT-5.5 or Opus 4.8 leads every row. The honest read is an open-weights coding agent one step below the closed frontier, and this one you can run on your own hardware.
Kimi-K2.7-Code hardware requirements
The system requirement to check is memory, and a 1T-parameter model needs a lot of it: count RAM and VRAM together. The GGUF builds below come from unsloth/Kimi-K2.7-Code-GGUF, each sharded into files of roughly 48 GB; the sizes listed are the total per build.
| Memory | Build to pick | File size |
|---|---|---|
| 320 GB | UD-IQ1_M | 303.9 GB |
| 384 GB | UD-Q2_K_XL | 339.5 GB |
| 512 GB | UD-Q3_K_XL | 463.9 GB |
| 640 GB | UD-Q4_K_XL | 583.7 GB |
| 768 GB and up | UD-Q8_K_XL | 594.5 GB |
When two builds both fit, take the larger one, and leave headroom for the KV cache: the 256K context is not free. If the format is new to you, start with what GGUF is.
How to run Kimi-K2.7-Code in Atomic Chat
Atomic Chat is a free local app for macOS, Windows and Linux. It includes a Hugging Face model browser and a built-in chat, with no manual llama.cpp build required.
- Download Atomic Chat for your platform and open it.
- Search for Kimi-K2.7-Code in the model browser and open Download Options.
- Pick the build that fits the memory you have, then start a chat.
The rest of the family is at every Kimi model you can run locally, including the general-purpose sibling Kimi-K2-Instruct-0905.
Kimi-K2.7-Code license
Both the code repository and the model weights are released under Moonshot's Modified MIT License, listed on Hugging Face as license "other". The MIT base permits use, modification and redistribution, including commercial use; the "modified" part is Moonshot's own added terms, so read the LICENSE file in the repository before you build a product on it.