Kimi K2.7-Code, developed by Moonshot AI, is an agentic, coding-focused large language model designed for long-horizon software engineering tasks. It is distributed with open weights on Hugging Face under a Modified MIT license and is also accessible via the Kimi API and Kimi Code platform.

Architecturally, K2.7-Code is a Mixture-of-Experts (MoE) model, comprising a total of $1$ trillion parameters with $32$ billion parameters activated per token. Its MoE design incorporates $384$ experts, from which $8$ are selected per token, and $1$ is shared across tokens. The model features $61$ layers, including a single dense layer. For attention mechanisms, it utilizes Multi-Query Attention (MLA), and its feed-forward network employs the SwiGLU activation function. Additionally, it integrates a MoonViT vision encoder, contributing $400$ million parameters, to support image and video input. The model employs native INT4 quantization for efficient deployment. It boasts a substantial context window of $256,000$ tokens ($262,144$). Operational constraints include a mandatory "thinking mode" for its agentic process, which cannot be disabled via API, and fixed sampling parameters (temperature $1.0$, top\_p $0.95$, n $1$, penalties $0.0$). The default maximum output length is $32,768$ tokens. For self-hosting, compatibility is provided with vLLM, SGLang, and KTransformers, though the model's $595$ GB disk size positions it as a server-class deployment.

Moonshot AI's internal benchmarks indicate that K2.7-Code consistently outperforms its predecessor, K2.6, across various coding and agentic evaluation suites. Notably, it achieved a $21.8\%$ improvement on the Kimi Code Bench v2, increasing from $50.9$ to $62.0$. It also showed significant gains on Program Bench ($+11.0\%$), MLS Bench Lite ($+31.5\%$), Kimi Claw 24/7 Bench ($+9.3\%$), MCP Atlas ($+9.5\%$), and MCP Mark Verified ($+11.4\%$). In comparisons against frontier models, K2.7-Code surpassed Claude Opus 4.8 on MCP Mark Verified ($81.1$ vs. $76.4$) and approached GPT-5.5's performance on MLS Bench Lite ($35.1$ vs. $35.5$). These benchmarks were conducted using Kimi Code CLI for K2.7-Code, Codex xhigh for GPT-5.5, and Claude Code xhigh for Opus 4.8.

A key cost optimization highlighted is a reported $30\%$ reduction in "reasoning-token" usage compared to K2.6, which translates to lower output-token costs, faster interactive CLI sessions, and an extended effective context window in multi-step agentic workflows.

Targeted use cases include repo-scale refactors where the agent can read files, edit across modules, and iteratively debug; code review by analyzing pull request diffs; complex tool-use workflows integrating Model Context Protocol (MCP) for CI checks, ticket updates, and file edits; and long-context analysis combining text, image, and video inputs for comprehensive prompt understanding.

The Kimi API is OpenAI-compatible, requiring adherence to the fixed sampling parameters and strict preservation of reasoning_content in multi-turn tool calls. Its competitive API pricing is $0.95 / 1\text{M}$ tokens for cache-miss input and $4.00 / 1\text{M}$ tokens for output, with a cached input rate of $0.19 / 1\text{M}$.

Strengths include its open weights, consistent performance gains over K2.6, competitive API pricing relative to closed-source models, and a strong showing on the MCP Mark Verified benchmark. Weaknesses noted are the reliance on first-party benchmarks, the inability to disable "thinking mode," locked sampling controls, the requirement to preserve reasoning_content for tool use, and the substantial size of the model weights for self-hosting.