AI MODEL PROFILE

Kimi K3

The first open 3 trillion parameter class model, released under a custom licence that needs reading before you build on it.

Source review Last reviewed August 7, 2026 By Moonshot AI

The facts at a glance

ProviderMoonshot AI
TypeOpen weight multimodal model
Current version or tierKimi K3, released late July 2026
ModalityText, image, and video in, text out
AccessOpen weights under the bespoke Kimi K3 Licence, which is not an OSI approved licence
Context or limits1,048,576 token context, 2.8 trillion total parameters with 104 billion activated across 896 experts
Free optionYes, by downloading the weights from Hugging Face
Paid priceHosted API pricing not published at time of checking

Pricing last checked August 7, 2026. Prices change and vary by region. Check the official pricing page ↗

Assessment

What it does well

  • The model card reports 93.5 percent on GPQA Diamond, 88.3 on Terminal-Bench 2.1, and 91.2 on BrowseComp
  • Frontier scale weights you can actually download and inspect
  • Full million token context

Where it falls short

  • The Kimi K3 Licence is bespoke and not OSI approved, so redistribution and commercial terms need reading rather than assuming
  • Hosted API pricing was not published when we checked
  • Ships as quantisation aware trained weights, which adds a deployment consideration

Privacy and data control

Self hosting the weights keeps your data on your own infrastructure. If you use Moonshot's hosted service instead, its own terms apply and we have not verified them.

The longer read

Open weights is not the same as open source

Kimi K3 can be downloaded, but it ships under a licence Moonshot wrote itself rather than MIT or Apache. That is increasingly common and it is not automatically a problem, but it does mean the terms are specific to this model. Read them before a commercial deployment. If licence simplicity matters more to you than scale, DeepSeek-V4-Flash under plain MIT is the easier answer.

How it performs by outcome

We assess products against a job, not a leaderboard. These are the outcomes this entry is compared under.

Worth comparing against

Grok 4.5 Terminal and agentic coding work where token efficiency matters DeepSeek-V4-Flash Teams that need to run a capable model on their own infrastructure Qwen3.8-Max Multi day autonomous coding projects Mistral Medium 3.5 Coding agents running on a small GPU cluster

How this profile was produced

Source review. Every fact here is drawn from the provider documentation linked below and checked on the date shown. We have not yet run this entry through a controlled test, and no scoring is implied.

Read the full review methodology →

Primary sources