Kimi-K2.7-Code

Updated
03.07.2026
Tools
Thinking
Vision
Reasoning
Code
Run

Run Kimi-K2.7-Code, a 1T-parameter coding model, locally with Atomic Chat. Offline agentic coding, fully private. Download free.

huggingface-cli download moonshotai/Kimi-K2.7-Code
from transformers import AutoModel
model = AutoModel.from_pretrained("moonshotai/Kimi-K2.7-Code")
More models
View all

  
  

NameSize / UsageContextInput
Kimi-K2-Instruct
Agentic tasks, coding128KText
Kimi-K2-Instruct-0905
256KText
At a glance
  • License: Other
  • Context length: 256K tokens
  • Languages: Multilingual
  • Minimum hardware: ~592 GB VRAM
  • Strengths: reasoning, coding and on-device inference
Overview
Kimi-K2.7-Code is a coding-focused large language model from Moonshot AI, part of the Kimi K2 family.
  • License: Other
  • Context length: 256K tokens
  • Languages: Multilingual
  • Minimum hardware: ~592 GB VRAM
  • Strengths: reasoning, coding and on-device inference
Highlights
  • +21.8% on Kimi Code Bench v2 — over the previous K2.6, using ~30% fewer thinking tokens.
  • Agentic tool calling — drives multi-step workflows across a full production stack.
  • Code across 10+ languages — backend, infrastructure, frontend, and ML/data engineering.
  • Vision + 256K context — reads large files, screenshots, and diagrams in one pass.
What it is good at
Moonshot tuned this version for end-to-end coding agents, and its capabilities reflect that. It reports +21.8% on Kimi Code Bench v2 over the previous K2.6, uses about 30% fewer thinking tokens, and beats Claude Opus 4.8 on the MCP Mark Verified agent benchmark.
  • Agentic tool calling — it drives multi-step workflows, calling tools and chaining actions to finish a task rather than answering in one shot.
  • Code reasoning and generation — it works across 10+ languages and a full production stack, from backend services and infrastructure to frontend and ML/data engineering.
  • Long-context vision — with vision and a 256K window, it can read large files, screenshots, or diagrams and reason over them in a single pass.
Running it locally
This is a heavy model. The full-precision weights run well past 600GB, so most local setups use a quantized build. A 2-bit quant lands around 350GB and needs roughly that much combined system RAM and VRAM; people run it on a single 24GB GPU with 256GB RAM via CPU offloading at low token rates, or on multi-GPU rigs for faster output. The 262144-token context adds further memory on top during long sessions.
huggingface-cli download moonshotai/Kimi-K2.7-Code
You can serve the weights with vLLM or llama.cpp, or skip the setup and load Kimi-K2.7-Code in Atomic Chat with one click, which handles the download and RAM offloading for you.
License
Kimi-K2.7-Code is released by Moonshot AI under a Modified MIT License (listed here as "other"). It permits free use, modification, and redistribution of the weights, including for commercial projects, with the terms of that license applying.
Desktop



macOS
(M1 or better)
Download
Windows
(x64)
Download
Linux
(x86_64)
Download
Frequently asked questions
Kimi-K2.7-Code is an open-weight coding model from Moonshot AI in the Kimi K2 family. It is a Mixture-of-Experts model with about 1058.6B total parameters and a 262144-token context window, tuned for agentic software engineering tasks like multi-step code generation and tool use. Moonshot reports it beats Claude Opus 4.8 on the MCP Mark Verified agent benchmark.
This is a large model, so plan for a lot of memory. A 2-bit quantized build is around 350GB and needs roughly that much combined system RAM and VRAM. People run it on a single 24GB GPU paired with 256GB of system RAM using CPU offloading at a few tokens per second, while multi-GPU rigs run it faster.
Yes. The weights are published openly on Hugging Face under a Modified MIT License, so you can download and run them at no cost. The license permits free use, modification, and redistribution, including commercial use, under its terms. Running it locally in Atomic Chat is free; the only cost is the hardware to host it.
Yes. Once you download the weights, the model runs fully on-device with no internet connection. Loading it in Atomic Chat keeps every prompt and response on your own machine, so your code stays private. You only need a connection for the initial download.
It is built for coding and agentic work: multi-step code generation, tool calling, and reasoning across large codebases in 10+ languages. It reports +21.8% on Kimi Code Bench v2 over K2.6 while using about 30% fewer thinking tokens. For general writing and conversation, Moonshot recommends the more well-rounded K2.6 instead.