Kimi K3, the Moonshot AI model in numbers

Kimi K3 at a glance

Kimi K3 is a language model from Moonshot AI, released on July 16, 2026. That makes it 3 months old as of October 22, 2026. Its weights are published: you can download it and run it on your own servers, or use one of the 82 API providers that list it on models.dev. It has a reasoning mode: it can "think" before answering, which helps with multi-step problems but uses more tokens, and therefore more money.

We file it under open-weight models (68 models in this category on the site).

Benchmark results

According to results published by Epoch AI (checked on October 22, 2026), Kimi K3 has a capability index of 80.7 out of 100, computed on 5 benchmarks from our basket.

That puts it at #25 in our weekly model ranking.

• GPQA Diamond (PhD-level science questions): 93.1% — the best tracked model, Gemini 4 Argon, reaches 96.4%
• OTIS Mock AIME (competition mathematics problems): 97.2% — the best tracked model, GPT-5.5 Pro, reaches 100%
• FrontierMath (very hard research-level mathematics): 72.2% — the best tracked model, GPT-6 Astra, reaches 93.7%
• FrontierCode (advanced programming tasks): 44.2% — the best tracked model, Claude Opus 5.5, reaches 54.6%

What it can do

Its context window reaches 1,048,576 tokens, roughly 1,573 pages of text at 500 words per page: enough to load a whole book or a mid-sized codebase. A single answer can run to 1,048,576 tokens.

It accepts as input: text, image and video.

It can call tools (functions, APIs) and return structured JSON output, which makes it a possible foundation for an AI agent.

The weights are published on Hugging Face (moonshotai/Kimi-K3). See also our list of open-source models.

How much does Kimi K3 cost?

Kimi K3 costs $3 per million input tokens and $15 per million output tokens (developer's price, checked on October 22, 2026).

To put it in perspective: reading 20,000 tokens (about thirty pages) and writing 2,000 costs roughly $0.09; a thousand of those tasks, $90.

That is 15 times the median output price of open-weight models ($1).

Tokens read back from cache cost $0.30 per million, which helps an agent that keeps resending the same context.

To estimate your own budget, use the agent cost calculator.

When to choose it

Given its category, Kimi K3 is mainly aimed at projects where you want control over your data or hosting, or want to adapt the model to your field.

Its FrontierCode result (44.2%, above the median of tracked models) makes it a serious candidate for coding agents.

Its very large context window makes it worth considering for analysing large sets of documents in a single request.

Before adopting it, test it on your own examples: benchmarks measure standard tasks, not your job.

Comparable models

In the same category, the models closest to Kimi K3 are:

• GLM-5.3 (Zhipu AI (Z.ai), index 78.8) — see the comparison
• Kimi K2.5 (Moonshot AI, index 78.1) — see the comparison
• GLM-5 (Zhipu AI (Z.ai), index 77.2) — see the comparison
• Kimi K2.6 (Moonshot AI, index 84.4) — see the comparison