01Overview02Models03Rankings04Compare05Quiz06Guide

Model

Kimi K2.7 NO.38

Open-weight coding refresh of Kimi K2 — stronger agentic tool use at low cost.

CompareBack to the index

Moonshot AI · Kimi · Open · Open weights

Specification

Fig.01 — Kimi — Moonshot AI

CHINA

Context window
262Ktokens
Max output
33Ktokens
Input price
$0.95/ 1M
Output price
$4/ 1M
Cached input
$0.19/ 1M
Throughput
—
Class
Open
Modalities
Text, Image, Video
Weights
Open
Parameters
1000B
License
Modified MIT
Knowledge cutoff
Apr 2025
Released
Jun 2026

Overview

Kimi K2.7 (shipped as K2.7-Code) is Moonshot's June-2026 coding-focused refresh of K2.6 — the same 1T-parameter open-weight MoE (32B active, Modified MIT) with forced thinking, retuned for agentic software engineering and MCP tool workflows while using roughly 30% fewer thinking tokens. Moonshot reports gains over K2.6 across its coding suites and an MCP-tool score edging Opus 4.8, at roughly one-seventh the token cost. Independent third-party benchmarks were not yet available at launch, so its coding claims are vendor-reported — validate on your own tasks; the Claude/GPT frontier still leads proven end-to-end reliability.

Strengths

  • Among the strongest open-weight agentic coders and MCP tool users
  • Roughly one-seventh the token cost of the closed frontier
  • 256K context; 1T MoE with only 32B active
  • Modified MIT open weights, self-hostable
  • ~30% more token-efficient thinking than K2.6

Trade-offs

  • Only vendor benchmarks published so far (no independent runs)
  • Long-horizon reliability still trails Claude/GPT
  • Large 1T MoE — serving cost and latency
  • 32K output cap

Fit

Best for

  • Cost-efficient agentic coding within a supervised loop
  • MCP and tool-use workflows
  • Self-hosted agent stacks
  • Long-context repository work

Not ideal for

  • Unsupervised mission-critical autonomous coding
  • Workloads needing independently-verified benchmarks

Capabilities

Ten axes, normalized 0–100 and scored the same way across the whole catalog — so a 78 here means what a 78 means anywhere else on the bench.

Fig.02 — Capability radar

0 — 100

Reasoning
84
Coding
78
Math
86
Writing
82
Knowledge
84
Speed
55
Agentic
78
Vision
55
Multilingual
80
Long Context
84

Benchmarks

Public results, with independent third-party runs marked. Bars normalize percentages against 100 and Elo ratings against a 1500 ceiling.

No public benchmark scores are recorded for this model.

Alternatives

Models in roughly the same class — the ones worth weighing against this record.

Compare side by side
01Kimi K2.6Moonshot AIOpen-weight agentic powerhouse with long context and video input.02GLM-5.2Zhipu AIOpen-weight coding flagship at a fraction of frontier cost.03DeepSeek-V4 ProDeepSeekOpen MIT flagship — 1.6T-parameter MoE with native vision and a 1M-token context at throwaway prices.

Index

PreviousGLM-4.7NextKimi K2.6
All models/EDUCATION