Kimi K3

Jul 2026
Kimi K3
Compare in
on
$1.15per task?
$3 / $15per 1M in / outcache $0.30
98 tok/sfirst token in 1.1 s?
1Mcontext

Rankingat low effort?

Full ranking →
1GPT-6 Astra96.32Gemini 3.8 Flash95.33Grok 4.694.94Gemini 3.1 Pro Preview94.14Muse Spark 1.394.1
···
87Gemini 2.5 Pro84.488Kimi K384.288Qwen3.6 27B84.2

Strong at

A
Long context#52 of 273Knowledge#55 of 318Vision#30 of 124
B
Coding#34 of 132Agents & work tasks#58 of 175Reasoning#100 of 334Science#88 of 312

Alternatives?

ModelScoreCostPer taskWhat changes
CheaperGemini 3.8 Flashmedium+11%-19%*$0.93Sampling parameters are ignored+1BetterGPT-5.6 Lunahigh+6%-96%$0.044OpenQwen3.8 2.4T A95Bthinking+11%+88%$2.16Context: 1M → 262K+6CheaperGPT-6 Astralow+11%-29%$0.82Monitoring can return HTTP 403+2CheaperGrok 4.5high+11%-23%$0.88Context: 1M → 500K

Run it yourself

estimate
Fits on✗ 8 GB (phone, laptop)✗ 16 GB✗ 24 GB GPU✗ 48 GB GPU✗ 80 GB GPU
Licencekimi-k3commercial use with conditions

History

5 versions
7580859095Jul 25Oct 25Jan 26Apr 26Jul 26Release dateGPQA DiamondK2 · JulK2 · SepK2.5K2.6K3
7580859095Jul 25Jan 26Jul 26Release dateGPQA DiamondK2 · JulK2 · SepK2.5K2.6K3

Benchmarks

in Chat & Knowledge
AA Intelligence Index 4.3.230.1#56
AA Omniscience Index3.9#55
AA Omniscience · Accuracy45.8#33
AA Omniscience · Hallucination Rate77.1#154
Humanity's Last Exam25#100
GPQA Diamond84.2#88
AA-LCR (long context) 1.179.3#52
CritPt (physics)3.1#91
GDP.pdf · AA all-pass18.6#27
ARC-AGI-1 · public evaluation75.8#38
ARC-AGI-1 · semi-private65.7#41
ARC-AGI-2 · public evaluation16.1#38
ARC-AGI-2 · semi-private12.4#43
GPQA Diamond · Epoch84.8#56
OTIS Mock AIME 2024 202568.9#77
Chess Puzzles18#55
Mystery Game Puzzles12#50
HLE fulllab43.5
MathArena Apexlab65.6
AA Cyber Index successeslab41

Watch out 4

Sampling parameters are fixed›

temperature, top_p, n and penalty values are fixed; omit them from requests. Source ↗

No public image URLs›

Vision input rejects public image URLs; use base64 or ms:// file references in a content array. Source ↗

Web search is being updated›

Web search is not recommended for near-term production workflows; avoid depending on it in production. Source ↗

Thinking mode is always enabled›

K3 always has thinking enabled; control effort with the top-level reasoning_effort field using low, high or max. Source ↗

FAQ

How much does Kimi K3 cost?›

$3 per 1M input tokens and $15 per 1M output tokens; cached input $0.30 per 1M; cheapest via Relace at $0.80 / $3. A typical benchmark task costs about $1.15 at low effort.

Is Kimi K3 free to use?›

There is no free API; the cheapest paid price is $0.80 per 1M input tokens. The weights are open, so you can run it yourself.

How good is Kimi K3, and what is it best at?›

#100 of 334 in Chat & Knowledge, #34 of 132 in Coding, #36 of 133 in Agents & Tool Use, #17 of 195 in Writing & Language, #30 of 124 in Vision & Documents. Strongest at: Long context, Knowledge, Vision.

Is Kimi K3 better than Kimi K2.6?›

On GPQA Diamond, Kimi K3 scores 93.5 vs 91.1 — higher. It is 43% more expensive per task.

How can I use Kimi K3?›

Through the API of AkashML, Alibaba, BaseTen, Chutes, DeepInfra and 13 more.

What is the context window of Kimi K3?›

1M tokens, with up to 944K output tokens.

When was Kimi K3 released? Is there a newer version?›

Released 2026-07-16. It is the current version in its line.

Is Kimi K3 open source?›

Yes, the weights are open (kimi-k3 licence).

Can I run Kimi K3 locally?›

The weights are open and can be self-hosted; hardware needs are not confirmed yet.

Can Kimi K3 read images?›

It accepts text and images, video as input.

What is Kimi K3?›

Kimi K3 is an open-weights language model from Moonshot AI, released Jul 2026, the successor to Kimi K2.6.

What are the best alternatives to Kimi K3?›

Gemini 3.8 Flash (cheaper), GPT-5.6 Luna (better), Qwen3.8 2.4T A95B (open weights).