Explore ProductsAI WorkspaceApps & SystemsManaged agentsWhite labelAPI
PlatformsDesktop AppChrome ExtensionWeb App
Explore ResourcesResourcesUse casesCase studies
Mission
SPIRITT logoSPIRITTMoonshot AIKimi K3

Kimi K3 is open frontier intelligence

Released July 16, 2026 by Moonshot AI, Kimi K3 is a 2.8-trillion-parameter sparse MoE with a 1M-token context and native text/image/video understanding. Independent Artificial Analysis puts it near the closed frontier, and Arena ranked it #1 for frontend code. On SPIRITT you run it inside a full agent workspace, not only a chat box.

Try Kimi K3 Now
See how it works

Across X on Kimi K3

Hands-on power users, open-weights takes, Frontend Arena hype, and skeptical vibes — not just Moonshot PR.

Official launch card

From Moonshot: Kimi K3 is open frontier intelligence built for long-horizon coding, knowledge work, and multimodal agents.

What Moonshot built it for

A sparse MoE at 2.8T total parameters with only a small expert set active per token, native multimodality, and a 1M-token context for repository-scale and long document work.

Positioned as downloadable open weights (Kimi K3 License / commercial-use conditions apply) plus hosted API and Kimi Code. Always-on reasoning with effort controls in the product stack.

Open weights2.8T MoE1M contextNative multimodalCoding agentsKimi Code

AA Intelligence Index

Near frontier
Claude Fable 5
60
GPT-5.6 Sol (max)
59
Kimi K3 (max)
57
Claude Opus 4.8 (max)
56
Grok 4.5 (high)
54
Gemini 3.6 Flash
50

Source: Artificial Analysis around K3 launch. Kimi K3 ~57 sits just behind Claude Fable 5 and GPT-5.6 Sol class leaders.

Where Kimi K3 is strong vs other models

Same score means nothing alone. These are head-to-head charts so you can see where it leads, where it is close, and what that means for real agent work.

Frontend coding
Winner

Arena Frontend Code Elo

Kimi K3
1679
Claude Fable 5
1631
GPT-5.6 Sol
1618

Blind human preference. Kimi K3 debuted #1 ahead of Claude Fable 5.

Coding index
Near frontier

AA Coding Index (reported)

Kimi K3
76.2
Typical mid pack (ref)
60

Independent AA coding composite reported around launch (~76.2).

Terminal agents
Near frontier

Terminal-Bench 2.1 (reported)

GPT-5.6 Sol
88.8
Kimi K3
88.3
Gemini 3.6 Flash (Meta harness ref)
78.9

Launch coverage puts K3 near GPT-5.6 Sol on Terminal-Bench 2.1 (~88.3 vs ~88.8).

Deep agentic SE
Winner

DeepSWE (reported)

Kimi K3
67.5
Muse Spark 1.2 (Meta harness)
59.3

AI Release Tracker / launch tables list K3 DeepSWE around 67.5%.

Efficiency

Cost per AA Intelligence task

Kimi K3
$0.94
GPT-5.6 Sol
$1.04
Opus 4.8 (ref)
$1.80

AA: K3 ~$0.94 per task, similar to Sol (~$1.04), cheaper than Opus-class (~$1.80).

API price
Near frontier

Input price ($ / M tokens)

Kimi K3
$3.00
Typical frontier list (ref)
$10.00

Hosted API list around launch: $3 / $15 per M input/output, cache ~$0.30.

Context
Winner

Context window (tokens)

Kimi K3
1000000
Typical 200K closed model
200000

1,048,576-token context for long repos and multimodal traces.

Limitations

Hallucination rate (AA, %)

Better closed peers (ref)
30%
Kimi K3 (AA)
51%

AA noted elevated hallucination (~51%) despite strong accuracy. Test factual workflows carefully.

Sources: Moonshot Kimi K3 launch post; Artificial Analysis Kimi K3 profile and launch-day writeups; Arena Frontend Code leaderboard; Vals AI index posts. Some coding board figures are launch-coverage composites—re-check live leaderboards before production bets. On SPIRITT you run K3 with tools, browser, and durable memory.

How It Works

From zero to Kimi K3 running real work in a cloud agent environment

01

Open a workspace

Open a workspace and land in a fully equipped cloud computer: browser, files, terminal, integrations, and memory. No local setup. No thin chat box pretending to be an agent.

Open a SPIRITT workspace
02

Pick Kimi K3

Open the model picker and choose Kimi K3. Same model shipped for agentic work, now inside a workspace that already has tools, browser control, and durable context.

Select Kimi K3 in the model picker
03

Build or automate

Tell it what to ship or what to run. Kimi K3 can code, call tools, drive the browser, coordinate multi-step work, and keep going while you step away. The point is not another chat window. It is an agentic environment where Kimi K3 actually does the job.

Build and automate with Kimi K3

Try Kimi K3 in an agentic environment

A real cloud environment, model picker, tools, and memory for open frontier agent work.

Try Kimi K3 Now
Official Moonshot launch
Questions

Frequently asked questions

01What is Kimi K3?+
Kimi K3 is Moonshot AI's July 2026 flagship: a ~2.8T sparse MoE with 1M context, native multimodality, and open weights. It targets frontier coding and agent work while remaining downloadable under Moonshot's K3 license terms.
02How is Kimi K3 different from previous Kimi models?+
It jumps into the closed-frontier pack on independent AA (~57), took Frontend Code Arena #1, and ships as open weights at 3T-class scale with 1M context and native vision/video—not just a chat API bump.
03What is Kimi K3 best for?+
Long-horizon coding agents, frontend generation, tool-using knowledge work, and teams that want frontier-adjacent quality with open-weight deployment options.
04What are some of the main limitations of this model?+
Verbose always-on reasoning can burn tokens; AA flagged higher hallucination rates; self-hosting needs serious multi-accelerator hardware; license is open weights with commercial conditions, not unrestricted MIT.
05What is a good example of using Kimi K3 well?+
Open a SPIRITT workspace, pick Kimi K3, and hand it a multi-file product UI or agent workflow with browser and tools—where Arena-style frontend strength and long context actually show up.
06Where is Kimi K3 best to test for full agentic capabilities?+
SPIRITT Workspaces. Pick Kimi K3 in the model picker and run real work in a fully equipped cloud environment: tools, browser, files, terminal, memory, and durable sessions. That is where the model can act as an agent, not just chat.
Buy from builders who use what they sellBuilt usingSPIRITT