Overview
Kimi K2.7 Code is a specialized programming iteration of the K2 series released by Moonshot on 2026-06-12—not a major K3 version, aiming to continuously refine coding capabilities while waiting for K3. 1.1T parameters MoE / 32B activated / 256K context / native INT4 quantization deployment, and for the first time in the K2 series, it adds MoonViT 400M Vision Encoder, natively supporting Image/Video input.
Compared to K2.6, it shows significant improvements on three coding benchmarks: Kimi Code Bench v2 +21.8% / Program-Bench +11% / MLS Bench Lite +31.5%, with Agent autonomous execution tasks (Kimi Claw 24/7, MCP Atlas, MCP Mark Verified) averaging about 10% improvement. Average token consumption reduced by 30%—this time clearly targeting the "overthinking" issue.
What's special about Moonshot's release note: They did not pick a benchmark to claim surpassing GPT-5.5 or Opus 4.8, but frankly acknowledged the gap. The official public statement "GPT-5.5/Opus 4.8 about 70 points, K2.6 about 50 points, K2.7 Code 60+ points" is a rare pragmatic attitude among domestic large model release notes.
On 2026-09-22, Moonshot released Kimi Code Desktop 1.0, the official desktop client for Kimi Code, with macOS builds for both Apple silicon and Intel plus a Windows build available at the same time; downloads are at kimi.com/code.
Key Features
- 1.1T MoE / 32B Activated / Native INT4: MoE architecture with 32B activated, native INT4 quantization, significantly lower deployment cost compared to dense models of similar capability
- 256K Long Context + Long-Range Coding: Significantly improved long-context coding instruction following, enhanced long-range coding task performance, 13-hour continuous coding + 300 Agent cluster continuing from K2.6
- MoonViT Multimodal Expansion: First time adding 400M Vision Encoder to K2, natively supporting Image and Video input, can read screenshots/UI mockups/demo videos
- 30% Token Consumption Reduction: Targeted improvement of "overthinking" tendency, average token consumption reduced by about 30%, improved cost structure for heavy usage
- High-Speed Version (Available Next Monday): Approximately 180 t/s (median input) / 260 t/s (short context), 5-6 times faster than standard version, priced at 2x standard version
- Kimi Code CLI Official Harness: Officially recommended agent framework, forces preserve_thinking mode to retain multi-turn reasoning context
- Kimi Code Desktop 1.0 (2026-09-22): Official desktop client for Kimi Code, with macOS builds for both Apple silicon and Intel plus a Windows build available at the same time; downloads at kimi.com/code
Use Cases
- Cost-sensitive small and medium teams replacing Claude Code / Cursor for daily Coding Agent
- Private deployment scenarios (INT4 + open-source weights) for enterprise-level AI coding platforms
- Default upgrade for Kimi Code Plan subscribers (no switch needed, price unchanged)
- Full-stack development requiring native support for UI screenshots / demo videos for requirement understanding
- R&D scenarios requiring 13-hour level long-range coding (dependency updates, large-scale refactoring, document synchronization)
- Desktop-based coding workflows on macOS (Apple silicon or Intel) and Windows via Kimi Code Desktop 1.0
Pros
- All three coding benchmarks comprehensively surpass K2.6, Agentic capability +10%
- Token consumption reduced by 30%, improved cost structure for heavy usage
- Native INT4 + open-source weights, friendly for private deployment
- First addition of MoonViT multimodal encoder, UI/video input ready out of the box
- Official honest acknowledgment of gap with GPT-5.5/Opus 4.8, a rare pragmatic stance
- Official desktop client Kimi Code Desktop 1.0 available for macOS and Windows since 2026-09-22
Pricing
Open-source weights: Available for download on HuggingFace (moonshotai/Kimi-K2.7-Code), commercial use and self-hosting allowed.
Official API: Same price as K2.6 (input $0.22/million tokens, output $0.88/million tokens), Code Plan subscribers automatically upgraded to K2.7 Code.
High-speed version (available next Monday): Approximately 180-260 t/s, priced at 2x standard version.
Comparison with K3 timeline: The highlight this year remains Kimi K3, targeting to match GPT-5.5 / Opus 4.8.
Summary
Kimi K2.7 Code is the most noteworthy incremental iteration among domestic open-source coding models released on 2026-06-12—coding benchmarks +20-30%, Agentic +10%, Token reduction 30%, price unchanged, almost a no-brainer benefit for Kimi Code Plan subscribers. Its capability still lags behind GPT-5.5 / Opus 4.8 by one tier, but Moonshot's proactive public acknowledgment of the gap makes it more trustworthy than peers' "benchmark edge cases." If you are currently using K2.6 for coding, upgrade directly; if you are cost-sensitive on Cursor / Claude Code, K2.7 Code + Kimi Code CLI is one of the most cost-effective open-source alternatives today; if you pursue top-tier capability with sufficient budget, Claude Code (short-term wait-and-see due to event impact) or GPT-5.5 + Codex is still recommended. Since 2026-09-22, the official desktop client Kimi Code Desktop 1.0 has also been available for macOS and Windows.
Version History
- Kimi Code Desktop 1.0 ships for macOS and Windows (2026-09-22): Kimi (Moonshot AI) released Kimi Code Desktop 1.0, the official desktop client for Kimi Code, with macOS builds for both Apple silicon and Intel plus a Windows build available at the same time; downloads are at kimi.com/code.
- Kimi K2.7 Code (2026-06-12): 1.1T MoE / 32B activated / 256K / INT4 / MoonViT multimodal; coding benchmarks +20-30%, Agentic +10%, Token reduction 30%; same price as K2.6
- Kimi K2.6 (2026-04-20): 13-hour continuous coding + 300 Agent cluster, chosen as base for Cursor Composer 2.5
- Kimi K2 / K2.5 (2025-2026): First trillion-parameter MoE open-source release, establishing leading position in domestic coding models