Modelpeers
Kimi K3 character

ModelMoonshot AI · Joined Jul 16, 2026

Kimi K3

Flagship Model at Moonshot AI | Open-weight agentic coding and browsing | 1M-token context

Available for Agentic coding, web research and computer use

Compare
Official pageReport a correction

The basics

Checked Oct 22
Status
AvailableFull weights on Hugging Face under the Kimi K3 License since Jul 27, 2026.
Model ID
Price
$3 in, $15 outPer 1M tokens, since Jul 16
Context
1,048,576 tokens
Speed
17.4 tokens a second30-day median, first token in 3.82s

Updates

New release

Kimi K3.1 Pro and K3.1 Pro Highspeed released

Moonshot released Kimi K3.1 Pro, a post-trained version of Kimi K3 with a 1-million-token context window, plus a Highspeed variant with 2-3x the speed. Both are available on Kimi.com, Kimi Work, Kimi Code and the Kimi API, and full open weights are due by Nov 4, 2026.

Now available

Kimi K3 available in GitHub Copilot

Kimi K3, hosted by GitHub on Fireworks AI, rolled out to Copilot Pro, Pro+, Max, Business and Enterprise at $3 per 1M input tokens, $15 per 1M output tokens and $0.30 per 1M cached input tokens.

Full changelog, by month

Strengths and weaknesses

Strengths

  • Agentic coding

  • Web research

  • Computer use

  • Image and video understanding

  • Science

  • Law and government

  • Coding

Weaknesses

  • !

    Output becomes unstable if your harness drops earlier thinking

  • !

    May make unexpected decisions on your behalf, so set explicit limits

  • !

    Moonshot says the user experience trails Claude Fable 5 and GPT 5.6 Sol

  • !

    Thinking is always on and defaults to max effort

  • !

    Running shell commands reliably

Role

Open-weight model at Moonshot AI

  1. Kimi K3 character

    Kimi K3Current

    Jul 2026 to now · 4 months

Works with

From recorded changes
GitHub Copilot character
GitHub Copilot

Agent · 2 shared changes

Kimi app

App · 2 shared changes

Kimi API

API · 1 shared change

Similar models

GLM-5.3 character
GLM-5.3

$1.40 in, $4.40 out

DeepSeek-V4-Pro character
DeepSeek-V4-Pro

$1.32 in, $3.96 out

Claude Opus 5.5 character
Claude Opus 5.5

$4 in, $20 out

Similar models for now. Real viewing data comes after launch.

Cost vs peers

5th cheapest of 6 similar models by output price.

Standard API rates for uncached input on the shortest prompts, from each maker's pricing page, or from models.dev for models not researched here yet. DeepSeek-V4-Pro: Peak hours. Off-peak prices are half.

Cost and specs

Cost

Per 1M tokens

Input
$3.00
Output
$15.00
Cached input
$0.30

Specs

Context
1,048,576 tokens
Parameters
2.8T total, 104B active
Experts
896, with 16 selected per token
Vision encoder
MoonViT-V2, 401M parameters
Input
Text, images and video
Reasoning effort
Low, high and max (default). Thinking is always on.
License
Kimi K3 License

Where it works

PlacePrice per 1M tokens, in / outUptime
Kimi APIkimi-k3
Kimi.com, Kimi app, Kimi Work and Kimi CodeAvailable
GitHub CopilotPro, Pro+, Max, Business and Enterprise. Admins must enable it.
OpenRoutermoonshotai/kimi-k3
Vercel AI Gatewaymoonshotai/kimi-k3
Self-hostedvLLM, SGLang or TokenSpeed
InferenceNet$0.23 / $6.97100%
Sail Research$0.15 / $7.15100%

Prices and uptime are OpenRouter's figures for each provider's cheapest endpoint. Uptime covers the last 24 hours.

Benchmarks

Higher is better. Green is Kimi K3. Dark marks are similar models.

Independent

Self-reported by Moonshot AI

BrowseComp

Hard web browsing questions

91.2

DeepSWE

Long software engineering tasks

67.5

OSWorld-Verified

Computer use tasks

84.8

What people say

Reviews

No reviews yet. Be the first to review Kimi K3.

Used Kimi K3? Tell other builders what worked and what did not.

Incidents

1 incident, 2 h 54 min recorded in the last 90 days, from Moonshot AI's status page. Only major incidents or ones lasting 30 minutes or more are counted. Fewer records do not prove better uptime.

Incidents per week, last 13 weeks. Most in one week: 1.

Aug 3, 2026Last 7 days

Errors in GitHub Copilot for about three hours

Resolved

From about 09:20 to 12:14 UTC, Copilot users saw elevated errors with Kimi K3 because of an upstream provider issue. No other models were affected.

Writing habits

Creative writing · EQ-Bench, among 51 models tracked here

Slop score

9.7

Overused AI words and phrases. Less slop than 86% of tracked models.

Repetition

3.7

Repeated words and ideas. Less repetition than 52% of tracked models.

Story length

5,488

Average characters per story. Longer than 4% of tracked models.

Research sources

Researched Oct 22, 2026. Each section's own sources open with its Show sources button.