Kimi K2.7-Code released as open-weight coding model

Category: AI
Status: developing
Sources: Hugging Face model card, HN discussion
Summary: Moonshot AI released Kimi K2.7-Code on 2026-06-12 as an open-weight agentic coding model under a Modified MIT license. The model card describes a Mixture-of-Experts architecture with about 1 trillion total and 32 billion activated parameters, a 256K context window, and roughly a 30 percent reduction in thinking-token usage versus Kimi K2.6. HN discussion focused on cost: commenters noted Chinese open-weight models price far below Anthropic Opus while approaching its quality for many coding tasks, and several reported evaluating a move off proprietary agents to opencode or similar harnesses with K2.7.
Why it matters: Cheaper open-weight coding models with usable agent harnesses raise migration pressure on teams paying premium per-token rates for proprietary coding agents.
Follow-up: Track independent coding-benchmark results and real cost comparisons against Claude Code and Codex.