Back to the index
Moonshot AI · Kimi
No. 35

Kimi K2.6

Open-weight agentic powerhouse with long context and video input.

OpenOpen weightsText, Image, Video
Field notes
Context window
262Ktokens
Max output
66Ktokens
Input price
$0.95/ 1M tokens
Output price
$4/ 1M tokens
Cached input
$0.16/ 1M tokens
Modalities
Text, Image, Video
Knowledge cutoff
Jan 2026
Released
Apr 2026
Parameters
1000B
License
Modified MIT
Approx. speed
—
Overview

What this model is

Moonshot's Kimi K2.6 is a 1T-parameter open-weight MoE (32B active, Modified MIT) built for agentic coding and long-horizon autonomous work, adding native video input and an agent-swarm system over K2.5. It ties GPT-5.5 on SWE-bench-class coding at roughly 80% lower cost, with a 262K context. It is among the strongest open challengers, though real-world reliability still trails the Claude/GPT frontier.

Strengths

  • Top-tier open agentic and coding benchmarks
  • 262K context for large repositories
  • Native multimodal input including video
  • Roughly 80% cheaper than frontier closed models
  • Modified MIT open weights

Trade-offs

  • Long-horizon reliability below Claude/GPT
  • Large 1T MoE — latency and serving cost
  • Agent-swarm claims outpace consistent results
  • Multilingual weaker outside English/Chinese

Best for

  • Agentic coding within a supervised loop
  • Long-context repo understanding
  • Multimodal (image/video) analysis
  • Cost-efficient autonomous workflows
  • Self-hosted agent stacks

Not ideal for

  • Unsupervised mission-critical agentic coding
  • Ultra-low-latency needs
Capabilities

Capability profile

Normalized 0–100 scores, comparable across the whole catalog.

Reasoning84Coding76Math86Writing82Knowledge84Speed55Agentic76Vision60Multilingual80Long Context84
Reasoning
84
Coding
76
Math
86
Writing
82
Knowledge
84
Speed
55
Agentic
76
Vision
60
Multilingual
80
Long Context
84
Benchmarks

How it scores

Public benchmark results, with independent third-party results where available. Bars normalize percentages to 100 and Elo ratings to a 1500 ceiling.

Independent index
Artificial Analysis Intelligence Index
Composite of ~9–10 independent evals · Artificial Analysis
43/ 100
BenchmarkResult
Artificial Analysis Intelligence Indexindependent
GeneralArtificial Analysis
43%
SWE-bench Verifiedindependent
CodingSWE-bench
80.2%
SWE-bench Proindependent
CodingSWE-bench Pro
58.6%
GPQA Diamondindependent
ReasoningIndependent aggregators
90.5%
LiveBenchindependent
GeneralLiveBench
72.2%
DeepSWEindependent
CodingDatacurve
24%
LMArena Eloindependent
GeneralLMArena
1466Elo
Alternatives

Comparable models

Models in a roughly similar class — worth weighing against this one.

GLM-5.2Zhipu AI

Open-weight coding flagship at a fraction of frontier cost.

MiniMax M3MiniMax

Open-weight multimodal model with native video input and 1M context.

Kimi K2 ThinkingMoonshot AI

Deep-reasoning open variant of Kimi K2 for hard problems.