LLM Reference

Kimi K3

Released
2026-07-14
Last refreshed
2026-08-24
Status
Researched 4d ago
Open weightsCommercial use: conditionalMultimodalRAGAgentsLong contextVisionJSON / Tool useHighlight

Kimi K3 is worth evaluating for rag, agents, and long context when its provider route and context window match the workload.

Use it for

  • Teams evaluating rag, agents, and long context
  • Workloads that can use a 1.05m context window
  • Buyers comparing 1 tracked provider route

Do not use it for

  • Workloads where another current model has stronger sourced task evidence
Specifications
Family
Kimi
Released
2026-07-14
Context
1.05m
Max output
1,048,576
Parameters
2.8T total; 16 of 896 experts active
Architecture
Mixture of Experts
Specialization
general
Openness
Open weights
License
Kimi K3 LicenseCommercial use: conditional
Weights
Available
Code
Unknown
Created by

Lossless long-context AI innovation

Beijing, China
Founded 2023
Website
Pricing
Output / 1M
$15.00
Input / 1M
$3.00

Cheapest of 1 route · Moonshot AI Kimi · cache read $0.300

About

Kimi K3 is Moonshot AI's 2.8-trillion-parameter flagship multimodal model for long-horizon coding, knowledge work, deep reasoning, and agentic workflows. It uses Kimi Delta Attention, Attention Residuals, and a sparse MoE design (16 of 896 experts active), supports a 1,048,576-token context window, text/image/video input, always-on reasoning, ToolCalls, strict JSON Schema structured output, automatic context caching, and partial mode through Moonshot's OpenAI-compatible API.

Top use-case fit: coding, agents, and build tasks

RAG

Included by capability and metadata signals in the decision map.

Agents

Included by capability and metadata signals in the decision map.

Long context

Included by capability and metadata signals in the decision map.

Provider price ladder

Compare API pricing across 1 providers for input and output tokens, batch, and cached reads when available.

ProviderInput / 1MOutput / 1MCacheRoute
Moonshot AI Kimi$3.00$15.00read $0.300
Serverless

Capabilities

VisionMultimodalReasoningJSON / Tool useStructured OutputsPrompt Caching

Benchmark peer barsfor RAG

No task-mapped benchmark peers are available for this model yet.

Benchmark scores(13)

Scores are benchmark-specific and are direction-aware: the same numeric gap can mean very different outcomes across suites. Use the leaderboard context and this model's provider route to decide whether the winning margin is meaningful for your workload.
BenchmarkScoreVersionEvaluationSource
Program Bench77.8KimiCode harness; max reasoning effortObserved 2026-07-16Source
Terminal-Bench 2.188.3Terminal-Bench 2.1; KimiCode harness; max reasoning effortObserved 2026-07-16Source
Google-Proof Q&A93.5GPQA-Diamond; max reasoning effortObserved 2026-07-16Source
Humanity's Last Exam43.5HLE-Full without tools; max reasoning effortObserved 2026-07-16Source
Humanity's Last Exam — With Tools56.0HLE-Full with tools; max reasoning effortObserved 2026-07-16Source
BrowseComp91.2BrowseComp; 300K-token context compaction; max reasoning effortObserved 2026-07-16Source
DeepSearchQA95.0DeepSearchQA F1; max reasoning effortObserved 2026-07-16Source
Toolathlon73.2Toolathlon-Verified; max reasoning effortObserved 2026-07-16Source
MCP-Atlas84.2500-task public subset; 100-turn limit; Gemini 3.1 Pro judge; max reasoning effortObserved 2026-07-16Source
AutomationBench30.8600-task public subset; official GitHub setup; max reasoning effortObserved 2026-07-16Source
WorldVQA51.0WorldVQA ForceAnswer; max reasoning effort; five-run meanObserved 2026-07-16Source
MMMU Pro81.6MMMU-Pro; official protocol; three-run mean; max reasoning effortObserved 2026-07-16Source
CharXiv84.8CharXiv RQ; three-run mean; max reasoning effortObserved 2026-07-16Source

Migration checks

No linked migration route is available for this model yet.