Kimi K3 is Moonshot AI’s new 2.8-trillion-parameter flagship model. Moonshot says K3 is available through Kimi products and the Kimi API, and that the full model weights are scheduled for release by July 27, 2026.
The practical answer is: use Kimi.com, Kimi Work, Kimi Code, or the Kimi API if you need access now. If you want downloadable weights, use only Moonshot’s official release links when they appear, because similar-looking third-party uploads can be incomplete, repackaged, or unsafe.
Quick answer
Kimi K3 is a large multimodal reasoning model for coding, knowledge work, long-context tasks, and agent workflows. Moonshot’s official launch post says it has 2.8 trillion total parameters, native vision support, a 1-million-token context window, Kimi Delta Attention, Attention Residuals, and a sparse mixture-of-experts design.
It is not a small local model. Moonshot recommends supernode deployments with 64 or more accelerators for serving the full model. For most readers, the realistic starting point is the hosted Kimi app, Kimi Code, or the API rather than a workstation download.
Confirmed at a glance
| Question | Current answer |
|---|---|
| What is Kimi K3? | Moonshot AI’s 2.8T-parameter flagship model |
| Is hosted access live? | Yes, through Kimi.com, Kimi Work, Kimi Code, and the Kimi API |
| Open-weight date | Moonshot says full weights will be released by July 27, 2026 |
| Context window | Up to 1 million tokens, depending on product and request limits |
| Vision support | Yes, native visual understanding is listed |
| API model name | kimi-k3 |
| API pricing | $0.30 per million cache-hit input tokens, $3.00 per million cache-miss input tokens, $15.00 per million output tokens |
| Local laptop use | Not realistic for the full model |
Prices, download links, license terms, and rate limits can change. Check Moonshot’s API documentation and the official model repository before building anything permanent.
What changed on July 27?
The July 27 date matters because Moonshot tied it to the full model-weight release. The model itself was already available through hosted Kimi products before that date, but developers were waiting for the downloadable weights, technical report details, and serving support.
During this update, Hugging Face API checks from our environment timed out, so this page does not claim that a specific repository URL is live. The confirmed source is Moonshot’s own documentation, which says the full weights are to be released by July 27. If you are downloading, start from Moonshot’s official Kimi pages or Moonshot’s verified Hugging Face organization rather than a search result.
How to access Kimi K3 now
Moonshot lists four official access paths:
| Access path | Best for | What to check first |
|---|---|---|
| Kimi.com | Trying the model in a browser | Account availability and regional access |
| Kimi Work | Knowledge-work tasks and desktop workflows | Current desktop version and supported OS |
| Kimi Code | Terminal and IDE coding work | Login method, model selector, and project permissions |
| Kimi API | Building apps or running evaluations | Top-up requirement, pricing, rate limits, and request caps |
For API use, Moonshot’s docs show an OpenAI-compatible client setup using base_url="https://api.moonshot.ai/v1" and model="kimi-k3". That makes testing simple if you already use OpenAI-style SDK calls, but it does not remove the need to review data handling, billing, and limits.
Our guide to comparing AI tools is a useful checklist before switching production work to a newly released model.
Can you run Kimi K3 locally?
Not in the consumer sense. Kimi K3 is a sparse mixture-of-experts model with 2.8 trillion total parameters. Moonshot says it effectively activates 16 of 896 experts, uses MXFP4 weights with MXFP8 activations, and recommends supernode configurations with 64 or more accelerators for deployment.
That means a local laptop, gaming PC, or single workstation GPU is not the expected target for the full model. Smaller community conversions may appear, but they should be treated as separate artifacts until Moonshot or a trusted serving project confirms what they contain and how they perform.
What vLLM support means
The vLLM team says it has been preparing day-0 serving support for Kimi K3, including model integration, parser and cache work, Docker images, deployment recipes, and NVIDIA and AMD paths. That is important because a huge open-weight model is only useful if the serving stack can load it, route experts, cache long prefixes, and keep multimodal and tool-use behavior compatible.
This does not mean setup will be one command for normal users. It means the open-source serving path is being coordinated so serious infrastructure teams have a starting point when the weights and support files are available.
What remains unknown
Several details should be rechecked before relying on Kimi K3:
- The final official weight repository URL and file layout.
- The exact license attached to the released weights.
- Whether the technical report changes any benchmark interpretation.
- Real-world latency and cost under your own prompts.
- Whether third-party hosts are serving the official model or a modified copy.
- How Kimi K3 behaves in long-running agent sessions outside Moonshot’s recommended harnesses.
Moonshot also lists limitations, including sensitivity to thinking-history handling, excessive proactiveness on ambiguous tasks, and a remaining user-experience gap compared with the strongest proprietary models it names.
Safe download checklist
Before downloading or deploying any Kimi K3 weights:
- Start from Moonshot’s official Kimi blog, Kimi API docs, or verified organization profile.
- Check the repository owner, license, model card, file sizes, and recent commits.
- Avoid uploads that use similar names but are not linked by Moonshot or a trusted serving project.
- Verify checksums or file manifests when provided.
- Run initial tests with non-sensitive prompts and capped spending.
- Document where the model is hosted if your team handles private code, customer data, or regulated content.
This is basic caution for any high-profile model release. It matters more when demand is high and unofficial mirrors can appear quickly.
Official sources
Moonshot’s Kimi K3 technical launch post confirms the model size, architecture, availability paths, July 27 weight-release target, pricing, deployment recommendation, and limitations. The Kimi API K3 documentation confirms the kimi-k3 API model, access requirement, reasoning-effort behavior, 1M-context notes, limits, and billing model. The vLLM Kimi K3 preview explains the planned open-source serving support.
FAQ
Is Kimi K3 available?
Yes. Moonshot says Kimi K3 is available through Kimi.com, Kimi Work, Kimi Code, and the Kimi API.
Are the Kimi K3 open weights out?
Moonshot says the full weights are scheduled for release by July 27, 2026. Check Moonshot’s official links for the current download location before using any repository.
What is the Kimi K3 API model name?
The API documentation uses kimi-k3.
How much does Kimi K3 cost through the API?
Moonshot lists $0.30 per million cache-hit input tokens, $3.00 per million cache-miss input tokens, and $15.00 per million output tokens.
Can Kimi K3 run on a laptop?
The full model is not aimed at laptop use. Moonshot recommends large multi-accelerator serving setups, so most users should start with hosted access.
Is every Kimi K3 download safe?
No. Use official Moonshot links or trusted serving-project references, and avoid unverified mirrors or repackaged files.