Parameters
1.0T total / 32.0B active
MoE: total / active
Architecture
Mixture-of-Experts (MoE)
Released
11.06.2026
License
Modified MIT License (Kimi)
Input Modalities
Output Modalities
Context (native)
262,144 tokens
Context (extended)
262,144 tokens
About
Kimi K2.7 Code (moonshotai/Kimi-K2.7-Code) is Moonshot AI's coding-focused agentic model built upon Kimi K2.6 - 1T total parameters with 32B activated per token, released June 11, 2026 under the Modified MIT License. Architecture: Mixture-of-Experts with 61 layers (including 1 dense layer), attention hidden dimension 7168, MoE per-expert dimension 2048, 64 attention heads, 384 experts with 8 selected + 1 shared per token, Multi-head Latent Attention (MLA), SwiGLU activation, 160K vocabulary, a 400M-parameter MoonViT vision encoder, and a 256K-token context.
It delivers substantial improvements on real-world long-horizon coding tasks, strengthening end-to-end task completion across complex software engineering workflows, and improves token efficiency by reducing thinking-token usage approximately 30% compared with Kimi K2.6. The model ships with native INT4 quantization (the same method as Kimi-K2-Thinking), and forces thinking with preserve_thinking mode so reasoning context persists across tool calls.
Training Data Coding-focused agentic model built upon Kimi K2.6. Uses native INT4 quantization (same method as Kimi-K2-Thinking). Forces thinking and preserve_thinking mode.
Benchmark Scores
| Benchmark | Score | Date | Standing |
|---|---|---|---|
|
MLS-Bench-Lite
coding_agent
|
36.21%
|
23.08.2026 |
#7/9
better than 25%
|
|
Kimi Claw 24/7 Bench
agentic
|
40.40%
|
23.08.2026 |
#3/4
better than 33%
|
|
MCP-Atlas
general_agent
|
82.77%
|
23.08.2026 |
#13/54
better than 75%
|
|
MCPMark-Verified
agentic
|
41.29%
|
23.08.2026 |
#2/4
better than 67%
|
|
WildClawBench
coding_agent
|
62.35%
|
23.08.2026 |
#9/20
better than 58%
|
|
Kimi Code Bench V2
coding_agent
|
42.37%
|
23.08.2026 |
#5/7
better than 33%
|
|
Program Bench
coding_agent
|
69.06%
|
23.08.2026 |
#3/7
better than 67%
|
|
LHTB Solved
coding_agent
|
100.00%
|
23.08.2026 |
#1/1
only tracked model
|
What is a good MLS-Bench-Lite score? What is a good Kimi Claw 24/7 Bench score? What is a good MCP-Atlas score?
Hardware Requirements
Estimated VRAM per quantization at 4,096 tokens context — MoE: all experts must fit in VRAM, only 32.0B are active per token
| Quantization | Weights | KV Cache | Total VRAM | On your hardware |
|---|---|---|---|---|
| FP16 — 16-bit | 2000.0 GB | 5.0 GB | 2105.0 GB | Which GPUs can run it? Calculate on your GPU |
| INT8 — 8-bit | 1000.0 GB | 2.5 GB | 1052.5 GB | Which GPUs can run it? Calculate on your GPU |
| INT4 — 4-bit | 500.0 GB | 1.2 GB | 526.2 GB | Which GPUs can run it? Calculate on your GPU |
| FP4 — 4-bit | 500.0 GB | 1.2 GB | 526.2 GB | Which GPUs can run it? Calculate on your GPU |
Only datacenter GPUs — FP16 needs 2105.0 GB. For consumer hardware, choose the INT4 build with a multi-GPU setup.
Architecture
- Attention
- Multi-head Latent Attention
- MoE
- 384 experts · top-8 per token
- Layers
- 61
- Hidden size
- 7168
- Context
- 262K tokens
- RoPE θ
- 50K
- Parameters
- 1000000M
- Active params
- 32000M
Source: Hugging Face config.json · KimiK25ForConditionalGeneration · model repo
Training Pipeline
-
1
other
Coding-Focused Fine-Tuning from Kimi K2.6
Kimi K2.7 Code is a coding-focused agentic model built upon Kimi K2.6. With substantial improvements on real-world long-horizon coding tasks, it strengthens end-to-end task completion across complex software engineering workflows while improving token efficiency, reducing thinking-token usage by approximately 30% compared with Kimi K2.6. Uses native INT4 quantization (same method as Kimi-K2-Thinking). Forces thinking and preserve_thinking mode.
Training & Evaluation Datasets
| Name | Role | Size | Modalities | Collection |
|---|---|---|---|---|
| Coding-focused post-training corpus (built on Kimi K2.6) | rl | — | — |
Linked Resources
Kimi Code CLI
https://www.kimi.com/code
Moonshot AI Homepage
https://www.moonshot.ai/
Kimi-K2.7-Code on HuggingFace
https://huggingface.co/moonshotai/Kimi-K2.7-Code
Moonshot AI on ModelScope
https://modelscope.cn/organization/moonshotai
Moonshot AI on HuggingFace
https://huggingface.co/moonshotai
Model Deployment Guide
https://huggingface.co/moonshotai/Kimi-K2.7-Code/blob/main/docs/deploy_guidance.md
Modified MIT License
https://huggingface.co/moonshotai/Kimi-K2.7-Code/tree/main/LICENSE
Third Party Notices
https://huggingface.co/moonshotai/Kimi-K2.7-Code/blob/main/THIRD_PARTY_NOTICES.md
Moonshot AI Platform (API)
https://platform.moonshot.ai/
Moonshot API Docs - K2 Thinking
https://platform.moonshot.ai/docs/guide/use-kimi-k2-thinking-model#complete-example
Kimi-K2-Thinking (INT4 quantization reference)
https://huggingface.co/moonshotai/Kimi-K2-Thinking#4-native-int4-quantization
Kimi Discord
https://discord.gg/TYU2fdJykW
ProgramBench
https://programbench.com/
MLS-Bench
https://mls-bench.com/
MCP-Atlas Leaderboard (Scale Labs)
https://labs.scale.com/leaderboard/mcp_atlas
MCPMark
https://mcpmark.ai/
LHTB Leaderboard
https://zli12321.github.io/LHTB/leaderboard.html
WildClawBench
https://internlm.github.io/WildClawBench
Moonshot's large visual-language model collection
https://huggingface.co/collections/moonshotai/kimi-k25
Trend Analysis
24h Change
+41.6%
7d Change
+59.2%
30d Change
+0.3%
Current
132,686
followers
+0.2%
downloads_all_time
+2.1%
downloads
+0.3%
likes
-0.1%
No comments yet
Be the first to share your thoughts!