Kimi K3
The first open-weight model in the 3-trillion-parameter class: 2.8T mixture-of-experts, 1M context, text, image and video input.
In short. Kimi K3 is a huge AI model from Moonshot AI that anyone may download, with about 2.8 trillion parameters. It understands text, images and video and is designed for long coding, research and office tasks. A newer version, Kimi K3.1 Pro, is now out, but K3 is still the base of that line.
Summary
Kimi K3 is Moonshot AI's flagship and, at release, the first open-weight model in the 3-trillion-parameter class. Moonshot introduced it through its API in mid-July 2026 and published the weights and technical report on 27 July 2026, together with the training-infrastructure tools MoonEP, FlashKDA and AgentEnv.
It is a mixture-of-experts model with about 2.8 trillion parameters, of which roughly 104 billion are active per token. It has 93 layers: 69 use Kimi Delta Attention, a linear-attention design, and 24 use gated multi-head latent attention. A router picks 16 of 896 experts for each token, and 2 shared experts run on every token. A 401-million-parameter vision encoder lets it read images, and Moonshot's card also describes video understanding. The context window is 1,048,576 tokens. Weights ship in MXFP4, a 4-bit format used in training, for a 1.56 TB download.
Moonshot says the new attention design and the Attention Residuals technique give about 2.5 times the scaling efficiency of Kimi K2. Benchmark tables on the model card are Moonshot's own and compare K3 with closed models run under different harnesses, so treat them as vendor claims. The Kimi K3 License is MIT-style with two scale conditions: hosting businesses above US$20 million of revenue need a deal, and very large products must credit the model.
Good for
- Running a frontier-class open model on your own multi-GPU infrastructure
- Long-horizon coding, research and office-style agent tasks
- Reading documents, charts and video together with tool use
- Studying a very large linear-attention mixture-of-experts design
Watch out for
- The 1.56 TB checkpoint needs data-center hardware, not a workstation
- The license asks large model-hosting businesses for a separate agreement
- Benchmarks are self-reported and use vendor-chosen harnesses
- Kimi K3.1 Pro is newer and is the better starting point for new projects
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the K3 variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| K3 MXFP4 | 2.8T 104B active | 6.44 TB | 3.22 TB | 1.72 TB | Multi-node cluster |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
An MIT-style license with two scale conditions. Most companies can use it commercially without asking, but large hosting businesses need a deal and very large products must credit the model. Kimi K3.1 Pro uses the same terms under its own name.
- Separate deal needed to sell hosted access above a threshold
- Attribution or model name must be shown
See also
- Tencent's Hy4 preview: 770B parameters, 1M context and Apache 2.0
- Hydrogen-1.0: a 753B open science model built on GLM-5.3
- Meituan's LongCat-2.5: 1.6T parameters, vision and the plain MIT license
- Giant open models now ship in 4-bit: what it means for memory planning
- Moonshot releases Kimi K3.1 Pro: 2.8T parameters, open weights
Frequently asked questions
- What is Kimi K3?
- Kimi K3 is a huge AI model from Moonshot AI that anyone may download, with about 2.8 trillion parameters. It understands text, images and video and is designed for long coding, research and office tasks. A newer version, Kimi K3.1 Pro, is now out, but K3 is still the base of that line.
- Can I use Kimi K3 commercially?
- Yes, with conditions. Kimi K3 is released under the Kimi K3 License, which allows commercial use but adds extra rules: Separate deal needed to sell hosted access above a threshold; Attribution or model name must be shown. Read the license text before you ship a product.
- Is Kimi K3 open source?
- Kimi K3 is open-weight, not open source in the strict sense. You can download the model files, but the Kimi K3 License adds conditions that the Open Source Initiative definition does not allow.
- Can I run Kimi K3 on my own computer?
- Not realistically. The smallest variant, K3, needs about 1.72 TB at 4-bit precision, more than a single GPU server holds. It is meant for clusters or hosted services.
- When was Kimi K3 released?
- Kimi K3 was released by Moonshot AI on July 27, 2026.
- How long a context window does Kimi K3 have?
- Kimi K3 supports a context window of up to 1M tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Kimi K3 Open Day: weights, technical report and key infrastructure released (Kimi, 27 Jul 2026) kimi.com
- Kimi-K3 model card (Hugging Face) huggingface.co
- Deploying Kimi K3 on AWS (AWS Machine Learning Blog) aws.amazon.com
- Kimi K3 License huggingface.co
Release date and facts last verified November 8, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Kimi K3.1 Pro Moonshot AI | 2.8T 104B active | 1M | Kimi K3 License (Conditional) | 1.72 TB Multi-node cluster | |
| Qwen3.8-Max Alibaba Qwen | 2.45T 95B active | 1M | Qwen3.8-Max License (Conditional) | 1.69 TB Multi-node cluster | |
| DeepSeek V4.1 Flash DeepSeek | 552B 16B active | 1M | MIT (Permissive) | 561 GB GPU server | |
| AREX-2 BAAI | 27.4B | 256K | Apache 2.0 (Permissive) | 19 GB Gaming GPU |