Kimi K2.7 Code
by Moonshot AI
Coding-specialised open-weight model built on K2.6: 1T total / 32B active MoE, 256K context, Modified MIT licence. Released 2026-06-12, about a month before the flagship Kimi K3 (2026-07-16).
Lineage (then → 2026 now)
K2 (2025) → K2.5 → K2.6 (2026-04-20, general agent model) → K2.7 Code (2026-06-12, coding specialist) → Kimi K3 (2026-07-16; 2.8T total / 104B active, 1M context, custom Kimi K3 licence). K2.7 Code is not superseded as a coding variant of the K2 line, but K3 is the current flagship and the API’s top model.
Facts (Hugging Face card, MarkTechPost)
- Purpose: stronger end-to-end completion on long-horizon software-engineering tasks than K2.6.
- Architecture: 1T total / 32B active; 384 experts, 8 selected per token; MLA attention; MoonViT vision encoder; 256K context; Modified MIT licence; weights ~595 GB.
- Release date: 2026-06-12 (MarkTechPost); the card itself states no date (Hugging Face org page shows the repo updated Jun 15).
- Efficiency: about 30% fewer thinking tokens than K2.6 (card).
Benchmarks (vendor-reported at launch; independent results were pending)
Kimi Code Bench v2: 62.0 (K2.6: 50.9, hence the “+21.8%” headline; GPT-5.5: 69.0; Claude Opus 4.8: 67.4). Program Bench 53.6 (+11.0% vs K2.6; Opus 4.8: 63.8 per the earlier card read). MLS Bench Lite 35.1 (+31.5%). MCP-Atlas 76.0 (Opus 4.8: 81.3 per the earlier read). The model trails the closed frontier on these tests.
Plans
| Model | Context | Notes |
|---|---|---|
| kimi-k2.7-code | 262,144 | standard usage-based API |
| kimi-k2.7-code-highspeed | 262,144 | faster variant at a higher rate |
| kimi-k2.6 | 262,144 | previous generation |
| kimi-k3 (for comparison) | 1,048,576 | flagship, priced higher |
Current prices: see the vendor pricing page https://platform.kimi.ai/docs/pricing/chat
What changed in 2026
This model did not exist in 2025; K2 Thinking-era Kimi (2025) has been through K2.5, K2.6, K2.7 Code and K3 within a year.
See also
Merged from laptop copy (2026-10-01)
Practitioner timeline from the owner’s earlier version of this note (video-summary coverage; not checked against primary sources, treat as unverified):
| Date | Event |
|---|---|
| 2026-06-12 | Released, tested as possibly the best open coding model |
| 2026-06-17 | Reported to beat Opus 4.8 and GPT-5.5, “really cheap” (contradicted by the vendor benchmark table above, where it trails both) |
| 2026-06-18 to 06-25 | Broad practitioner adoption; Obsidian and NotebookLM pairings |
| 2026-07-07 | Reported as the first open-weight model available in GitHub Copilot |
The GitHub Copilot inclusion would be the substantive milestone: the first open-weight model offered inside Microsoft’s mainstream coding product. Related: Kimi K2 series, GitHub Copilot, Qwen, GLM-5.2.
Sources
- https://huggingface.co/moonshotai/Kimi-K2.7-Code (accessed 2026-10-01)
- https://huggingface.co/moonshotai (org page, update dates; accessed 2026-10-01)
- https://platform.kimi.ai/docs/pricing/chat (accessed 2026-10-01)
- https://www.marktechpost.com/2026/06/12/moonshot-ai-releases-kimi-k2-7-code-a-coding-model-reporting-21-8-on-kimi-code-bench-v2-over-k2-6/ (accessed 2026-10-01)
- https://devops.com/moonshot-ais-kimi-k2-7-code-targets-token-efficiency-in-agentic-coding/ (search result, not fetched)
- https://en.wikipedia.org/wiki/Kimi_(chatbot) (lead for K2.6 date 2026-04-20 and K3 dates; accessed via search 2026-10-01)