Try Kimi at kimi.com →

Kimi K3
Moonshot's 2.8T Open Model

Kimi K3 is the 2.8-trillion-parameter, open-weight model that moved Kimi into the frontier tier in July 2026, with native vision and a 1M-token context. Here are its specs, benchmarks, prices and how it compares with the newer K3.1 Pro.

2.8T · 104B active 1M context Open weights
AaText ▣Image ▶Video Routerpicks 16 EXPERTS (896 in total, sample shown) Token 1 → text: 16 experts on Token 2 → image: 16 experts on Token 3 → video: 16 experts on 104Bactiveof 2.8T parameters CONTEXT1M tokens
OVERVIEW

Kimi K3 at a glance

2.8TParameters

About 104 billion active for each token.

1MToken context

Up from 256K on K2.6.

VisionText, image, video

Reads images, screenshots and video natively.

$3 / $15API per 1M tokens

Input and output; $0.30 cached input.

Status, November 2026: K3 has been succeeded by Kimi K3.1 Pro (23 October 2026), which is cheaper and stronger. K3 is still available in the API and Kimi Code, but most people should use K3.1 Pro.
SPECS

Kimi K3 specifications

SpecKimi K3
Released16 July 2026 (apps and API); open weights 27 July 2026
DeveloperMoonshot AI
ArchitectureMixture-of-Experts with hybrid attention (Kimi Delta Attention plus Gated MLA)
Parameters2.8 trillion total, about 104 billion active per token
Experts896 experts, 16 active per token
Context window1M tokens
Input / outputText, images and video in; text out
ReasoningAlways on, with low, high or max effort (default max)
Weights formatMXFP4 weights, about 1.4–1.6 TB on Hugging Face
LicenceCustom Kimi K3 License (open weights, commercial use allowed)
API model IDkimi-k3
API price$3 input, $15 output, $0.30 cached input per 1M tokens

From Moonshot's model card, technical report and API pricing page.

HOW

How Kimi K3 works

Mixture of Experts

K3 has 896 “expert” sub-networks. For each token, a router switches on just 16 of them, so only about 104B of its 2.8T parameters do the work. That keeps it affordable to run for its size.

Hybrid attention

It mixes Kimi Delta Attention, a fast linear attention first tested in Kimi Linear, with Gated MLA. This is what makes the 1M-token context practical.

Built-in vision

Images and video go into the same model as text, so K3 can read screenshots, charts, documents and video frames directly.

Efficiency: Moonshot's technical report says K3 gets about 2.5× more out of the same training compute than K2, and its weights ship in 4-bit MXFP4 format to cut memory use.
BENCHMARKS

Kimi K3 benchmarks

Moonshot-reported results

BenchmarkKimi K3Claude Fable 5GPT-5.6 Sol
FrontierSWELong real-world coding projects81.286.671.3
SWE-MarathonVery long coding tasks42.035.039.0
DeepSWESoftware engineering agent tasks67.570.073.0
Terminal-Bench 2.1Command-line tasks88.388.088.8
BrowseCompFinding hard facts on the web91.288.090.4
HLE-FullExpert questions, no tools / with tools43.5 / 56.053.3 / 63.044.5 / 58.0
GPQA DiamondGraduate-level science93.592.694.1

From the Kimi K3 model card, all at max effort. Harnesses differ between models, so treat close results as ties.

What the numbers say

  • 1
    Best at long jobs. K3 led SWE-Marathon and BrowseComp, tasks that run for a long time or need many searches.
  • 2
    Close to the top on coding. It beat GPT-5.6 Sol on FrontierSWE and matched the leaders on Terminal-Bench 2.1, but trailed Claude Fable 5 on FrontierSWE and DeepSWE.
  • 3
    Behind on expert reasoning. On Humanity's Last Exam without tools, K3 scored 43.5 against 53.3 for Fable 5.
K3 AT LAUNCH VS RIVALS Kimi K3 Claude Fable 5 GPT-5.6 Sol FrontierSWE 81.2 86.6 71.3 SWE-Marathon 42.0 ★ 35.0 39.0 BrowseComp 91.2 ★ 88.0 90.4 HLE-Full (no tools) 43.5 53.3 44.5 Moonshot-reported, max effort, July 2026. ★ = K3 leads.

K3 led on long coding tasks and web research, but trailed on expert reasoning.

Independent results

TestKimi K3Note
LMArena Frontend Code Arena#1 at launch, 1,679 pointsBlind human votes on front-end code
Artificial Analysis Intelligence Index44 (max effort)K3.1 Pro later scored 55
Vals AI SWE-bench run93.4%Independent run, behind Claude Opus 5 at 97.0%
Artificial Analysis speedAbout 38–55 tokens/sSlow for a top model
Artificial Analysis cost per task$2.00Writes many tokens per task (about 160M on the index)

Third-party tests run by the named organisations. Our full take is in the Kimi AI review.

COMPARE

Kimi K3 vs K3.1 Pro

Kimi K3Kimi K3.1 Pro
Released16 Jul 202623 Oct 2026
Base modelKimi K3K3 with much more post-training
Context window1M tokens1M tokens
API input / output$3 / $15$2 / $10
Cached input$0.30$0.20
Artificial Analysis index4455
Kimi Code Bench 3.0 Long-horizon55.95%61.2%
Claim-level hallucination (Moonshot stress test)23.4%8.3%
Faster versionNoK3.1 Pro Highspeed

Choose K3.1 Pro for…

Almost everything: it's cheaper per token, stronger at coding and finance, and much less likely to make things up.

Stay on K3 if…

You have a tested pipeline that depends on K3's exact behaviour, or you use the k3-256k option in Kimi Code to save quota.

Switching is easy

On the API, change the model from kimi-k3 to kimi-k3.1-pro. Nothing else changes.

ACCESS

How to use Kimi K3

Kimi app

K3.1 Pro is now the default. K3 powers the extra-long chat mode with up to 1M tokens on the Max and Ultra plans.

Kimi Code

Choose k3 (1M context) or k3-256k. The 1M version uses about twice the quota of k3-256k.

Kimi API

Use the model ID kimi-k3 at https://api.moonshot.ai/v1. Batch jobs are billed at 60% of the list price.

Self-host

Download the open weights from moonshotai/Kimi-K3 on Hugging Face. You'll need a multi-GPU cluster.

import os from openai import OpenAI client = OpenAI( api_key=os.environ["MOONSHOT_API_KEY"], base_url="https://api.moonshot.ai/v1", ) reply = client.chat.completions.create( model="kimi-k3", messages=[{"role": "user", "content": "Explain this stack trace and suggest a fix: ..."}], reasoning_effort="high", # low | high | max (default max) ) print(reply.choices[0].message.content)

Keys and billing: Kimi API guide. Coding setup: Kimi Code guide. Plans: Kimi pricing.

WEIGHTS

Open weights and self-hosting

  • 1
    Where: moonshotai/Kimi-K3 on Hugging Face, with the 47-page technical report on Moonshot's GitHub.
  • 2
    Licence: the custom Kimi K3 License. Commercial use is allowed; reported terms add obligations only for products with over 100 million monthly users or $20 million in monthly revenue. Read the licence before you ship.
  • 3
    Hardware: the MXFP4 weights take about 1.4–1.6 TB, so you need a multi-node GPU cluster. NVIDIA Blackwell and AMD MI400 GPUs handle the 4-bit format natively.
  • 4
    Software: vLLM and SGLang publish Kimi K3 recipes. Check the current versions, as support was still settling after launch.
Download
# Open weights (about 1.4–1.6 TB). Accept the licence on Hugging Face first. pip install -U "huggingface_hub[cli]" huggingface-cli download moonshotai/Kimi-K3 --local-dir ./kimi-k3 # Serve with vLLM or SGLang using their Kimi K3 recipes, # on a multi-GPU cluster with enough memory for the full model.
LIMITS

Kimi K3 limitations

Slow and wordy

Artificial Analysis measured about 38–55 tokens per second and very long outputs, so tasks take a while and cost more than the token price suggests.

More made-up claims than K3.1 Pro

On Moonshot's own hallucination stress test, K3 made claim-level errors 23.4% of the time, against 8.3% for K3.1 Pro. Check facts.

Busy at launch

Heavy demand led to “engine overload” incidents in August 2026, and Kimi paused new paid sign-ups from 20 July until late September.

Hard to self-host

The weights alone need about 1.4 TB or more of GPU memory, far beyond any single workstation.

FAQ

Kimi K3 FAQ

Kimi K3 is a large language model from Moonshot AI, released on 16 July 2026. It's a Mixture-of-Experts model with 2.8 trillion parameters, about 104 billion active per token, native image and video understanding and a 1M-token context window. Its weights are open.

Kimi K3 is open-weight. Moonshot published the weights on Hugging Face on 27 July 2026 under a custom Kimi K3 License, not a standard open-source licence. It allows commercial use, with extra obligations only for very large products.

Kimi K3 costs $3 per million input tokens, $15 per million output tokens and $0.30 per million cached input tokens on the Kimi API. Kimi K3.1 Pro is cheaper at $2 and $10.

K3.1 Pro is K3 with much more training. It's cheaper ($2/$10 vs $3/$15), scores higher on the Artificial Analysis index (55 vs 44), makes far fewer factual errors on Moonshot's stress test and has a faster Highspeed version. Both have a 1M-token context.

Not on a normal computer. The weights are about 1.4–1.6 TB, so you need a multi-GPU cluster with well over a terabyte of GPU memory. For most people, the Kimi app or the API is the practical option.

Yes. K3 is still on the Kimi API as kimi-k3 and in Kimi Code as k3 and k3-256k, and it powers the extra-long 1M chat on Max and Ultra plans. K3.1 Pro has replaced it as the default model.

Yes. K3 has native vision, so it can read images, screenshots, charts and video as well as text. It answers in text.

Strong. It topped LMArena's Frontend Code Arena at launch and led Moonshot's tables on long coding tasks such as SWE-Marathon. For new projects, K3.1 Pro scores higher on Moonshot's coding tests.

Try Kimi's newest model

K3.1 Pro builds on K3 and is the default in the Kimi apps, with limited free use.