Kimi-K2.7-Code — Moonshot ships a coding-tuned trillion-parameter MoE

Moonshot released Kimi-K2.7-Code around June 12 — a coding-focused agentic model built on Kimi K2.6, same 1T-total / 32B-active MoE architecture (384 experts, 8+1 active, DeepSeek-V3-style MLA backbone), 256K context, under a Modified-MIT license. It targets real-world long-horizon coding and tool use. Like K2.6, the size puts it firmly in hosted / multi-GPU territory, so it changes no local planner pick — it is a cloud comparator the way DeepSeek V4-Pro and Command A+ are.

Verdict: A coding-specialized 1T-MoE built on K2.6 — frontier-grade, but hosted / big-iron, not a local pick

The take

The facts, verified against the Hugging Face model card and the vLLM recipe: Kimi-K2.7-Code is a coding-and-agentic variant built on top of Kimi K2.6, sharing the K2.5/K2.6 architecture — a 1T-total / 32B-active mixture-of-experts (384 experts, 8 routed + 1 shared per token) on a DeepSeek-V3 backbone with MLA attention, 256K context. Moonshot publishes it under a Modified-MIT license (the same attribution-carrying MIT variant as the existing Kimi K2.6 entry). The pitch is "substantial improvements on real-world long-horizon coding tasks" over K2.6; vendor-card benchmarks put it competitive with the closed frontier on coding/agent suites (Kimi Code Bench v2, MCP Atlas / MCP Mark Verified), though none of that is independently benchmarked yet — treat as a vendor claim.

Why it matters: the open-weight frontier for coding keeps getting sharper, and it is still Chinese-led (Kimi, DeepSeek, GLM, Qwen). A coding-specialized 1T MoE under a near-permissive license is a real option for teams that can host it — but "host it" is the operative phrase. 1T total parameters means datacenter-class hardware or a serious multi-GPU server even at aggressive quantization; this is not something that fits a single consumer rig.

Our call: no planner-pick change. Kimi-K2.7-Code sits in the same bucket as the existing Kimi K2.6 entry — frontier, hosted / big-iron, tracked as a comparator, not a local recommendation. If you already run K2.6 on rented or owned multi-GPU hardware for agentic coding, the coding-tuned 2.7 is the obvious upgrade to evaluate. For everyone running locally on a single box, your coding picks are unchanged: Qwen3-Coder-30B-A3B and Qwen 3.5 35B-A3B at 24 GB+, GLM-5.1 if you can host it.

Where this fits

Models: Kimi K2.6 · GLM-5.1 · DeepSeek V4-Pro · Qwen3-Coder-30B-A3B

Hardware: NVIDIA DGX Spark · Dual RTX 5090 · Mac Studio M3 Ultra 96 GB

Sources

Next step

Try this in the planner→