Moonshot AI describes Kimi K3 as the world's first open 3T-class model, but the complete weights and final license are not yet public as of July 21, 2026. Moonshot says the weights will be released by July 27 alongside additional technical information.
That makes the accurate short answer: Kimi K3 is an announced open-weight model whose complete release still needs to be verified. The official API and Kimi products are already available; API availability is not the same as downloadable weights.
Status checked July 21, 2026. Do not download files from unofficial “Kimi K3” sites or make commercial-use claims until Moonshot publishes the official repository, checksums, and license.
Open source, open weights, and API access are different
These terms are often mixed together in search results.
| Term | What it means |
|---|---|
| API available | A provider runs the model and accepts remote requests |
| Open weights | Developers can obtain the trained parameter files under a license |
| Open source | The release provides sufficient code and rights to inspect, modify, and redistribute according to an open-source definition |
| Technical report | Architecture, training, and evaluation details are documented |
K3's API is available today. Moonshot has announced complete weights. Whether the total package should be described as fully open source depends on the actual license, code, and release artifacts.
Kimi K3 release timeline
July 16, 2026: model and API launch
Moonshot introduced K3 as a 2.8T-parameter model with native visual understanding, a 1M-token context window, and long-horizon coding and knowledge-work capabilities. K3 became available through Kimi, Kimi Work, Kimi Code, and the official API.
By July 27, 2026: announced weight release
Moonshot says full model weights will be released by July 27. It also says more architecture, training, and evaluation details will arrive with the technical report.
After release, verify:
- the official repository owner;
- weight file names and checksums;
- tokenizer and configuration;
- inference code or supported runtimes;
- model license;
- commercial-use conditions;
- derivative and redistribution rules;
- supported quantization formats.
Can you download Kimi K3 now?
The official launch page does not provide complete public weights as of this article's verification date. Search results already contain unofficial domains using Kimi K3's name. A polished download page is not proof that a file is official or safe.
Use only links published by Moonshot AI, Kimi's official documentation, or a verified Moonshot organization on a recognized model repository. Check large-file hashes and read the license before loading third-party model artifacts.
Can Kimi K3 run locally?
Not in the ordinary meaning of “locally.” A laptop, Mac, workstation, or single GPU cannot host the full 2.8T model.
K3 uses a sparse Mixture-of-Experts design: 16 of 896 experts are activated for a token. Sparse activation reduces active compute relative to a dense 2.8T model, but all required experts and shared components still need to be stored and served across an infrastructure-scale system.
Moonshot recommends supernode deployments with 64 or more accelerators. That recommendation implies high-bandwidth networking, expert parallelism, distributed storage, static-shape execution, monitoring, and specialized inference engineering.
Approximate weight-storage scale
K3 uses quantization-aware training with MXFP4 weights. A simple four-bits-per-parameter multiplication gives a theoretical raw weight payload of roughly:
2.8 trillion parameters × 4 bits ≈ 1.4 terabytes
This is an illustrative lower-level estimate, not a deployment specification. Real storage and memory requirements include metadata, scales, non-quantized components, routing data, runtime buffers, KV cache, redundancy, and framework overhead. The official files will determine the actual footprint.
Hardware needed to self-host Kimi K3
Consumer GPU
Not realistic for the complete model. A quantized community derivative would still face expert storage and execution requirements far beyond a typical desktop.
Multi-GPU workstation
Also not a realistic full-model target. A workstation lacks the accelerator count, aggregate memory, and interconnect expected by K3's deployment design.
Multi-node cluster
Technically plausible only with sufficient accelerator memory, high-bandwidth networking, compatible kernels, and a runtime that supports KDA and extreme expert parallelism.
Supernode deployment
This is the official recommended direction: 64 or more accelerators in a large communication domain. Exact supported hardware, throughput, and memory numbers should come from the technical report and inference-maintainer documentation.
Kimi Delta Attention and inference support
K3's Kimi Delta Attention creates new prefix-cache requirements compared with conventional attention implementations. Moonshot says it has contributed a KDA prefill-cache implementation to the vLLM community, planned alongside the model release.
Do not assume an ordinary vLLM installation can load and efficiently serve K3 until the relevant release, version, kernels, and hardware support are documented.
API versus self-hosting
| Decision area | Official API | Self-hosting |
|---|---|---|
| Startup time | Immediate after account setup | Potentially substantial |
| Infrastructure | Managed | 64+ accelerator-class deployment direction |
| Scaling | Provider managed | Your responsibility |
| Data control | Subject to provider terms | Greater infrastructure control |
| Model updates | Provider managed | Manual testing and rollout |
| Cost model | Per token | Hardware, utilization, people, power, operations |
| Customization | API and prompting limits | Depends on license and release tooling |
API access is the practical route for most teams. Self-hosting may make sense for organizations with sustained traffic, strict data-control requirements, and existing large-model infrastructure.
What license questions matter?
When Moonshot publishes the license, check:
- Is commercial use allowed?
- Can a hosted API be built from the weights?
- Are modified weights and fine-tunes permitted?
- Can derivatives be redistributed?
- Are there user-count, revenue, or compute thresholds?
- Are there acceptable-use restrictions?
- Must “Kimi” or “Moonshot” attribution be retained?
- Do different files use different licenses?
Do not infer K3's license from an earlier Kimi release.
How to download Kimi K3 safely after release
- Start from Moonshot's official launch or documentation page.
- Confirm the repository belongs to the verified organization.
- Read the license before downloading terabytes of data.
- Compare checksums where provided.
- Review model configuration and custom runtime code.
- Run inference code in a restricted environment before granting credentials or network access.
- Verify that the selected runtime version explicitly supports K3.
- Document the exact model revision used in production.
Is Kimi K3 really open source?
The answer should be updated after July 27. If Moonshot publishes weights, permissive code, a clear license, and reproducible inference components, “open source” may be a reasonable broad description. If only weights are released under a restricted license, “open-weight” will be more precise.
Until then, the safest wording is that Kimi K3 is an announced open 3T-class model with complete weights scheduled for release.
Read the Kimi K3 architecture guide for KDA, Attention Residuals, MoE, and quantization. The Kimi K3 review covers capabilities and API access.
Frequently asked questions
Is Kimi K3 fully open source?
The final answer depends on the weight, code, and license release. As of July 21, Moonshot has announced complete weights by July 27, but the full release is not yet available for verification.
Where can I download Kimi K3?
Wait for an official Moonshot AI or Kimi link. Avoid unofficial domains and unverified files.
Can Kimi K3 run on a Mac?
The complete 2.8T model is not a realistic Mac deployment. It is designed for large distributed accelerator infrastructure.
How much VRAM does Kimi K3 need?
Moonshot has not published one simple VRAM number for a complete supported deployment. The official direction is a supernode with 64 or more accelerators.
Can Kimi K3 be used commercially?
Commercial weight use depends on the license that accompanies the release. API use is governed separately by Moonshot's API terms.