Model · Moonshot AI · Code
Kimi K2.7 Code
Moonshot's coding-agent model built on Kimi K2.6 — a 1T-parameter mixture-of-experts (32B active) tuned for long, real-world software tasks, using about 30% fewer thinking tokens than K2.6. Open weights, but sized for a multi-GPU server.
- Modality
- Code
- License
- Modified MIT (Open weights)
- Parameter size
- 1T total, 32B active (MoE, 384 experts, 8 selected per token)
- Context window
- 262,144 tokens 256K per the model card; Moonshot's own tests ran at 262,144 tokens. Weights ship in native INT4. Image and video input via a 400M MoonViT encoder (video only on Moonshot's API for now).
- Released
- June 12, 2026
- Last verified
- October 28, 2026
- Runs locally
- Yes
- Also handles
- text, vision, reasoning
What you need to run it
- Mac
- Yes, but slow. Any Apple Silicon Mac (M1 or newer) with at least 512 GB of memory. It runs, but much slower than on an NVIDIA card.
- Windows / Linux PC
- Server only. Needs a server with several graphics cards, about 640 GB (1128 GB recommended) of video memory in total — not a desktop PC.
- No graphics card
- Not practical. Too slow without a graphics card or Apple Silicon.
Only a top-spec 512 GB Mac Studio (Ultra chip) can load it, and that is tight; otherwise it needs a multi-GPU server.
Numbers are for: 3-bit community MLX build of the 1T model (about 471 GB); the maker's INT4 weights are about 595 GB. Our estimate from the model’s size — the maker publishes no figure.
Strengths
- Moonshot reports clear gains over K2.6 on its own coding and agent tests — Kimi Code Bench v2 62.0 against 50.9, MCP Atlas 76.0 against 69.4
- About 30% fewer thinking tokens than K2.6 for the same work, which cuts cost and wait time
- Same architecture as K2.5 / K2.6, so existing vLLM, SGLang and KTransformers setups carry over
- OpenAI- and Anthropic-compatible API from Moonshot for teams that don't want to host it
Weaknesses
- Needs a multi-GPU server to self-host — for a small business the realistic route is Moonshot's API or a hosted provider
- Licence is Modified MIT: products with more than 100M monthly users or more than $20M monthly revenue must display "Kimi K2.7 Code" in their interface
- Benchmarks are Moonshot's own and mostly in-house; it still trails GPT-5.5 and Claude Opus 4.8 on the same tables
- Thinking is always on — there is no fast instant mode
Try it
| Where | Type | Notes |
|---|---|---|
| Hugging Face | weights | Modified MIT; native INT4 |
| Moonshot AI platform | hosted-api | API key required; OpenAI- and Anthropic-compatible |
| Kimi Code | product | Moonshot's coding product |
Version history
- Kimi K3 Jul 2026
- Kimi K2.7 Code Jun 2026 Current
Official sources
- Model card model
- Licence docs
Change log
- — Added hardware requirements (Mac support, memory, NVIDIA). Source: size of the Hugging Face weights and mlx-community/Kimi-K2.7-Code-mlx-DQ3_K_M-q8.
- — Initial entry. Added 2026-10-28 from the popularity sweep: 1,425 Hugging Face likes and 90.6k downloads/month. Release date 2026-06-12 is from press coverage of Moonshot's launch — the model card gives no date and the Hugging Face repo was created 2026-06-11. Modified MIT licence (branding rule above 100M monthly users or $20M monthly revenue), 256K context; API pricing not recorded here.
Esc