# Kimi K3: Moonshot AI's open-weight 2.8T multimodal model > Kimi K3 is an open-weight mixture-of-experts model with 2.8 trillion total parameters, 104 billion activated per token, native vision and a 1M-token context window. The repository is an announcement artifact: weights, a licence and a technical report, with no inference code. **MoonshotAI/Kimi-K3** — Open Frontier Intelligence - Repository: https://github.com/MoonshotAI/Kimi-K3 - Stars: 8,878 · Forks: 748 - Language: Unknown - License: NOASSERTION - Published: 2026-09-18 · Updated: 2026-09-18 · Language: en - Canonical page: https://hysenlabs.com/projects/moonshotai-kimi-k3 ## What the README claims, in its own terms Kimi K3 arrives with strong self-description: an open-weight, native multimodal agentic model, the most capable model Moonshot AI has released to date, and what the README calls the world's first open 3T-class model. The headline numbers are a 2.8 trillion parameter total built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), native vision capabilities, and a 1-million-token context window. The stated ambition covers long-horizon coding, knowledge work and reasoning. The framing that matters for adopters is the word weights. This repository does not ship a product; it ships the model's weights under the Kimi K3 License, a technical report, and documentation of what the model is claimed to do. Everything else, hosting, serving and evaluation against your own tasks, is the adopter's job. ## The architecture in numbers The model summary table is the most concrete thing in the repository, and it is detailed enough to plan around. The model is a mixture-of-experts architecture with 2.8 trillion total parameters and 104 billion activated per token, spread across 93 layers, of which only one is dense. Attention is split between 69 KDA layers and 24 Gated MLA layers, with a 7168 attention hidden dimension and 96 heads. The expert structure is equally specific: 896 experts with 16 selected per token plus 2 shared experts, a latent MoE dimension of 3584, and a per-expert hidden dimension of 3072. The vocabulary is 160K tokens, the context length is 1048576, and the activation function is listed as SiTU-GLU. The README adds that a Stable LatentMoE framework scales MoE sparsity, with an approximate 2.5 times improvement in overall scaling efficiency over Kimi K2, a claim made by the authors and elaborated in the technical report rather than proven in the README itself. ## What the long-horizon claims cover Two capability blocks get the most space. The first is long-horizon coding: the README says the model operates with minimal human oversight, sustains long engineering sessions, works across massive repositories, and orchestrates terminal tools, with examples ranging from GPU kernel optimization and compiler development to vision-in-the-loop game development, CAD and chip design. The second is agentic knowledge work: deep research with interactive visualizations, widgets and dashboards, and motion design and video editing, attributed to the native multimodal architecture. These are the vendor's own scenarios, stated without published numbers in the README. What makes them checkable rather than merely promotional is that the repository bundles the full technical report as k3_tech_report.pdf alongside a linked tech blog post, so a serious evaluator can read the methodology before believing any of it. The README also states that the model understands text, images and video within the same model, which is what native multimodality means here. ## Getting the model, and what this repository actually contains The repository tree is four entries: a LICENSE file, the README, an assets directory for images, and k3_tech_report.pdf. There is no inference code, no serving stack, no configuration templates and no published GitHub releases. The weights are distributed separately: the README links the moonshotai organisation on Hugging Face and on ModelScope, with the licence badge pointing at the licence file on the weights repository. For trying the model without deploying anything, the documented path is the hosted chat at kimi.com. For deploying it yourself, the README documents no commands: the README's own pointers are the technical report and the blog post, and from there adopters move to the weights repositories. Anyone accustomed to repositories that include transformers examples or quantized builds should calibrate expectations; this is an announcement and weights distribution, and the serving burden, which for a 104 billion activated parameter MoE is not small, sits entirely with the adopter. ## The licence question GitHub reports the licence as NOASSERTION, which means the repository ships a custom licence rather than a recognised template, and the README names it the Kimi K3 License. Open-weight does not mean unrestricted: the licence governs research, deployment and further use of the weights, and the README's own phrasing is that releasing the full weights makes frontier intelligence openly available for research, deployment and further innovation, which is the authors' description of their own terms, not a legal summary. The practical instruction is simple to state without giving legal advice: read the LICENSE file before building anything commercial on these weights, and do not assume permissions by analogy with Apache or MIT licensed code, because custom frontier-model licences commonly carry conditions that template licences do not. ## The comparison problem, stated honestly Two ways to use Kimi K3 exist, and they are different products in practice. The hosted chat at kimi.com requires no infrastructure and gives you the model as a service, under the service's terms and limits. Self-hosting the weights from Hugging Face or ModelScope gives control over data, serving and fine-tuning, at the cost of infrastructure capable of running a 104 billion activated parameter MoE and a careful read of the custom licence. Against other open-weight frontier models, the honest position is that this repository gives you an architecture table and a technical report, not a benchmark harness or a deployment cookbook, so cross-model comparison work lands on the adopter. For teams whose decision rests on measured behaviour on their own tasks, the report is the starting document, and the weights are the way to verify it. ## Conclusion Kimi K3 fits research teams studying frontier-scale architectures, since the technical report and the full parameter breakdown are both here, and teams with serious serving infrastructure that want open weights for long-context multimodal work. It does not fit hobbyists expecting a runnable repository, and it does not fit commercial deployments that have not first read the custom Kimi K3 License. Before adopting, read k3_tech_report.pdf for the evidence behind the scaling and capability claims, check the licence terms against your use case, and confirm your infrastructure plan against the 104 billion activated parameters and 896-expert structure. The last push was on 2026-08-06, and the weights live on Hugging Face and ModelScope rather than in this repository. ## FAQ ### What is Kimi K3 good at? Per the README, its stated strengths are long-horizon coding with terminal tools, agentic knowledge work such as deep research with visualizations and dashboards, and native multimodal understanding of text, images and video within a 1M-token context window. ### How big is Kimi K3? The model summary lists 2.8 trillion total parameters with 104 billion activated per token, 896 experts with 16 selected plus 2 shared, and a 1,048,576-token context length. ### Does the Kimi-K3 repository include code? No. The repository carries the README, the licence, image assets and the k3_tech_report.pdf technical report; weights are distributed through Hugging Face and ModelScope, and no inference code is included. ## Sources - [Issues](https://github.com/MoonshotAI/Kimi-K3/issues) - [MoonshotAI/Kimi-K3 on GitHub](https://github.com/MoonshotAI/Kimi-K3) - [README](https://github.com/MoonshotAI/Kimi-K3/blob/main/README.md) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/moonshotai-kimi-k3