Available on Julius

Moonshot AI: Kimi K3

Explore long-context research and reproducible analysis with Moonshot AI’s open-weight model.

Try Kimi K3 on Julius

Explore a worked example

Available on Pro, Max, Ultra, and Business. Compare plans

Reviewed October 13, 2026 · Sources

Cohort heatmap and retention curves showing customer-weighted retention of 56.7% in month one and 50% in month two. — Customer retention dashboardIllustrative output · Fictional dataView full size

Best suited to: Long-context research and code-based analysis
Published context: 1,048,576 tokens
Native inputs: Text, images, and video (provider capabilities)
Released: July 27, 2026

Kimi K3 pricing

Compare the provider’s API rates with access through Julius.

Fireworks standard serverless API

Input / 1M tokens: $3
Output / 1M tokens: $15

Per million tokens for the standard Fireworks endpoint, not a Julius credit rate. Fireworks Fast costs 50% more ($4.50 input / $22.50 output); Priority and regional endpoints also differ.

Check provider pricing

Use your Julius plan

Available on Pro, Max, Ultra, and Business. Credits cover the work performed with models and tools.

View Julius plans and credits

Prices in USD, reviewed October 13, 2026. API token charges are separate from Julius subscriptions and credits.

What is Kimi K3?

Kimi K3 is Moonshot AI’s open-weight model for extended coding, research, and knowledge work. The published model card describes a multimodal architecture and a one-million-token context window.

Kimi K3 is useful to evaluate when a task combines a large amount of source material with several reasoning steps. In Julius, specify what the model should investigate and what decisions it should leave to you.

Open weights

Moonshot publishes the weights under the Kimi K3 license. Running it yourself requires suitable infrastructure and compliance with that license.

Extended knowledge work

Moonshot’s model card describes long coding sessions and research with tools, alongside a large context window.

Build a retention table with the right denominator

Two customer cohorts have different sizes. How should their retention be combined?

Worked example · Fictional sample data · Expected results calculated from the CSVs

1. Input
Cohort sizes and active users
2. Check
Weight by original cohort size
3. Deliver
Combined customer retention

Expected result

At month 1, June retains 70% and July retains 50%. Combined retention is 56.7%, because July’s cohort is twice as large. At month 2, combined retention is 50.0%.

Combined month 1: 56.7%
Combined month 2: 50.0%

Month 1 retention · percent of original cohort

June: 70.0%
July: 50.0%
Combined: 56.7%

What the analysis needs to catch

An unweighted average would report 60% at month 1. Aggregate retained customers and original cohort sizes before dividing.

Make this deliverable

A cohort retention table and reproducible calculation steps.

Turn the retention analysis into an interactive dashboard with cohort filters using the AI website builder.

Try this example in Julius

The prompt includes the sample data. Review it in chat before sending. Results may vary; compare your answer with the calculations shown here.

View the sample output (PNG)View the workflow and chart (SVG)Download retention.csv Download all data and calculations

View the prompt and sample data

Build a retention matrix by cohort and month since acquisition. Divide active customers by the original cohort size. Calculate combined retention using cohort-size weights rather than a simple mean of percentages. Both cohorts have complete observations through month 2. Produce a reproducible analysis and a short explanation of the denominator.

retention.csv

cohort,month_index,cohort_size,active_customers
June,0,100,100
June,1,100,70
June,2,100,60
July,0,200,200
July,1,200,100
July,2,200,90

These expected results come from the downloadable Python calculations. This example illustrates a workflow; it is not a measured benchmark of Kimi K3.

Using Kimi K3 in Julius

Start with your task

The try buttons carry Kimi K3 into a new chat when your plan includes it. Example buttons also fill the composer with the sample data. Review the prompt, or replace it with your own files and question, before sending.

If the model is unavailable on your plan, Julius explains the access requirement. You can choose another option in the model menu.

Research a question Work through an analysis in code Create a report

Know the boundaries

The published model’s visual capabilities do not guarantee image or video support in every application. Julius’s current Kimi configuration is intended for text and tool-based workflows.
Set explicit boundaries for open-ended tasks. A useful prompt says what to investigate, what to produce, and when to ask for input.

Model details and official research

Go to the original documentation for specifications, release details, and evaluation methods.

Kimi K3 model card and benchmarks

Moonshot’s model architecture, evaluation setup, weights, and license.

Kimi K3 API pricing on Fireworks

Standard, Fast, and Priority serving options and their prices.

Kimi K3: frequently asked questions

Is Kimi K3 available on Julius?

Yes. It appears as Kimi K3 in the model menu. Available on Pro, Max, Ultra, and Business plans.

How much does Kimi K3 cost?

Fireworks standard serverless lists $3 per million input tokens and $15 per million output tokens. Per million tokens for the standard Fireworks endpoint, not a Julius credit rate. Fireworks Fast costs 50% more ($4.50 input / $22.50 output); Priority and regional endpoints also differ. Julius subscriptions and credits are separate.

Is Kimi K3 open source or open weight?

Moonshot releases Kimi K3’s weights under a dedicated Kimi K3 license. “Open weight” is the precise description here; inspect the license before deploying or redistributing it. Using Kimi in Julius does not require downloading the weights.