← Models

Kimi K2.7 Code

Moonshot AI · Kimi

Moonshot AI's thinking-only multimodal open-weight model for long-horizon coding agents and tool use.

Part of Kimi family · Other versions: Kimi K2.6 , Kimi K2.5
Type
multimodal
Context
262K tokens
Max Output
33K tokens
Status
current
Input
$0.95/1M tok
Output
$4/1M tok
API Access
Yes
License
Modified MIT
chinese reasoning coding agentic tool-use multimodal long-context open-weights thinking-only image-input video-input
Released June 2026 · Updated July 10, 2026

Overview

Freshness note: Model capabilities, limits, licensing, and pricing can change quickly. This profile is a point-in-time snapshot last verified on July 10, 2026.

Kimi K2.7 Code is Moonshot AI’s coding-focused successor to Kimi K2.6. Released on June 12, 2026, it targets long-horizon software engineering, tool use, repository-level execution, and coding-agent workflows while keeping official API access and public weights.

This is a specialist rather than a universal K2.6 replacement. Moonshot recommends K2.7 Code for coding and K2.6 for broader writing, analysis, and conversation. K2.7 Code always runs with thinking enabled; Kimi Code requests made with thinking disabled fall back to K2.6.

Capabilities

The model is built for end-to-end engineering work: understanding a repository, following multi-step instructions, editing across files, calling tools, and carrying reasoning across an extended agent session. It supports text, image, and video input with text output, although video input is currently experimental and limited to Moonshot’s official API.

Moonshot reports approximately 30% lower reasoning-token use than K2.6 alongside improved coding and agent benchmark results. Those figures are provider-run evidence; teams should still evaluate the model in their own agent harness, tool set, and repository mix.

Technical Details

Official anchors at this snapshot:

  • 1T total parameters and 32B active parameters in a Mixture-of-Experts architecture.
  • 262,144-token context window and a 32,768-token maximum-output configuration in Kimi Code’s integration guide.
  • 61 layers, 384 experts with eight selected per token, Multi-head Latent Attention, and a 400M-parameter MoonViT vision encoder.
  • Native INT4 quantization for the open-weight release.
  • Forced thinking and preserved reasoning across multi-turn coding sessions.
  • Official weights and code under Moonshot’s Modified MIT license.

The model can be served through vLLM, SGLang, KTransformers, and compatible quantized-model tooling. Its 1T-parameter footprint still makes it a serious cluster or rented-infrastructure deployment rather than a casual laptop model.

Pricing & Access

Moonshot’s current API pricing per 1M tokens is:

  • Cache-hit input: $0.19.
  • Cache-miss input: $0.95.
  • Output: $4.00.

The listed input price is the cache-miss rate. A high-speed route is also listed at 0.38cache−hitinput,0.38 cache-hit input, 1.90 cache-miss input, and $8.00 output. Access is available through the Kimi API, Kimi Code, official Hugging Face weights, and compatible self-managed runtimes.

The Modified MIT license adds one material condition: commercial products above 100 million monthly active users or $20 million in monthly revenue must prominently display “Kimi K2.7 Code” in the interface.

Best Use Cases

Choose Kimi K2.7 Code for coding agents, long-running refactors, multimodal debugging, MCP and tool-heavy engineering workflows, and open-weight deployments where the team can support very large-model infrastructure.

Use K2.6 instead for general-purpose assistant work or when non-thinking mode matters. Smaller Qwen, Mistral, or Gemma checkpoints remain more practical for local developer machines.

Comparisons

  • Kimi K2.6: Broader general-purpose multimodal model with optional thinking; K2.7 Code is the newer thinking-only coding specialist.
  • Qwen3.7 Max: Proprietary 1M-context hosted flagship; K2.7 Code offers public weights, multimodal input, and lower direct API prices.
  • MiniMax M3: Another very large multimodal open-weight coding model; K2.7 Code has the more permissive license, while M3 offers a 1M context.