1T total params, 32B active per token (MoE) - 384 experts, 8 selected + 1 shared, 61 layers, 7168 attention hidden dimension, 64 heads, MLA, SwiGLU activation, 160K vocabulary. Built on Kimi K2.6 as a coding/agentic specialization, not a general chat replacement.
-
Context and I/O: native 256K context (262,144 tokens via API); native vision and multimodal tool use (PNG/JPEG/WebP/GIF images, MP4/MOV/AVI video).
-
Reasoning and tools: always-on thinking mode; Moonshot claims ~30% fewer thinking tokens than K2.6. Native function/tool calling with MCP support.
Open weights under a Modified MIT license on HuggingFace at moonshotai/Kimi-K2.7-Code. The license requires products with >100M monthly active users or >$20M monthly revenue to display “Kimi K2.7 Code” prominently.
Cloud-only for most users. A 1T-parameter MoE needs server-class hardware; there is no published Unsloth GGUF or consumer-grade quantization, so self-hosting is not practical on a workstation or DGX Spark. Run it via the Kimi API, OpenRouter, or vLLM/SGLang on a cluster.
Cloud API: $0.95/1M input, $4.00/1M output, $0.19/1M cache-hit input. A kimi-k2.7-code-highspeed variant runs at ~180 tok/s (up to 260 tok/s in short contexts) for 2x the price.
Benchmarks (Moonshot self-reported): Kimi Code Bench v2 62.0, Program Bench 53.6, MLS Bench Lite 35.1, Kimi Claw 24/7 Bench 46.9, MCP Atlas 76.0, MCP Mark Verified 81.1.
AI-generated content marks
This model embeds text watermarks in generated text and adds C2PA provenance metadata to supported files such as .png, .jpg, and .svg.
Marks can be lost through editing, screenshots, or format conversion, so their absence does not prove a file is human-made.
Provider transparency docs →