Kimi K3 is open-weight, but you cannot download it yet. Moonshot has committed to releasing the full weights by July 27, 2026 on Hugging Face; until then you reach K3 through the app and API only. This guide covers when the weights land, where they will live, who hosts K3 now, and the hardware reality of running a 2.8-trillion-parameter model yourself.
Is Kimi K3 open source?
Yes, it is being released as an open-weight model, and at 2.8 trillion parameters it is being called the largest open-weight model shipped to date. "Open-weight" means Moonshot publishes the trained weights for you to download, run, and fine-tune, as it did for earlier Kimi flagships. It is not the same as fully open source: you get the weights, not necessarily the training data or pipeline.
The commitment is confirmed at the staff level, not just in marketing. Asked on X whether open-sourcing K3 would be "another DeepSeek moment" for the field, Moonshot's Xinyu Zhou (@zxytim) replied simply, "Will do."
The important caveat today is timing. As of this writing the weights are not on Moonshot's Hugging Face org, which still tops out at Kimi K2.7 Code. So K3 is announced as open-weight but is currently usable only as a hosted model.
When do the Kimi K3 weights drop?
Moonshot's own tech blog states the full weights will be released by July 27, 2026, alongside a technical report with the architecture and training details. Expect them under Moonshot's Hugging Face org, following the Modified MIT license it has used for recent Kimi releases; the final license ships with the files.
Until then, the model card, exact distribution formats, and the official active-parameter count are not public, so any "download Kimi K3" link before that date is not the real thing.
Where to run Kimi K3 today
You do not have to wait for the weights to use K3, you just cannot self-host it yet.
Live today is Kimi's own stack: the app at kimi.com, Kimi Code, and the API (model ID kimi-k3 at https://api.moonshot.ai/v1), billed at $3/$15 per million tokens. See the full pricing breakdown.
Expected once the weights ship: third-party hosts should light up quickly. Fireworks is a likely day-zero host, its Kimi partner page served K2.5, K2.6, and K2.7 at launch, and Baseten has opened a K3 waitlist rather than a live endpoint. Treat both as coming-soon, not available now. Multi-model aggregators tend to follow too; platforms like AIReiter already route Chinese open models such as DeepSeek V4 and GLM 5.2 through an Anthropic-compatible API.
For self-hosting, expect the usual local stack, Ollama, vLLM, and GGUF quantizations, to add K3 support after the weights drop, as they did for the K2 line.
The hardware reality of a 2.8T MoE
Kimi K3 is a Mixture-of-Experts model with 2.8 trillion total parameters (16 of 896 experts active per token), and Moonshot recommends serving it on supernodes with at least 64 accelerators. Self-hosting is not a single-GPU hobby.
For a concrete sense of scale, the previous-generation Kimi K2.7 Code, at 1 trillion parameters, needs roughly these resources (K3, at 2.8x the parameter count, will need substantially more):
| Precision | Approx. VRAM (K2.7 Code, 1T) |
|---|---|
| FP16 (full) | ~2,308 GB |
| INT4 | ~577 GB |
| Aggressive 2-bit (Unsloth dynamic) | ~325 GB |
Two things ease it. Moonshot ships MoE weights at native INT4 with BF16 attention, so a 4-bit quant is close to training precision, not a lossy afterthought; and Kimi Delta Attention gives up to 6.3x faster decoding in million-token contexts, which helps long agentic sessions. Even so, a realistic K3 self-host means a multi-GPU server or a quantized build on serious hardware, not a desktop.
Should you self-host or use a hosted model?
For almost everyone, hosted is the right answer. At $3/$15 per million tokens on the API, you would need very high, sustained volume before the cost of a multi-GPU rig (or a rented supernode) pays for itself against a managed endpoint. Self-hosting earns its keep in three cases: strict data residency or air-gapped requirements, fine-tuning K3 on your own data, or running at a scale where owned hardware genuinely undercuts per-token pricing.
If none of those apply, use the API or a host like Fireworks, and treat the open weights as insurance, the option to move in-house later, rather than a day-one plan.
FAQ
Is Kimi K3 open source?
It is open-weight: Moonshot will publish the downloadable weights, expected under a Modified MIT license. That is not the same as releasing training data, but it does allow self-hosting and fine-tuning.
When are the Kimi K3 weights released?
By July 27, 2026, per Moonshot's tech blog, on its Hugging Face org, alongside a technical report. They are not downloadable before then.
Can I run Kimi K3 on Ollama?
Not yet, because the weights are not public. Once they are released, expect Ollama, vLLM, and GGUF builds to add support as they did for the K2 series.
What hardware do I need to run Kimi K3?
A lot. Moonshot recommends at least 64 accelerators for serving. As a reference point, the 1T K2.7 Code needs roughly 577 GB of VRAM at INT4; K3 at 2.8 trillion parameters needs considerably more, so plan on a multi-GPU server or a quantized build on high-end hardware.
Where can I use Kimi K3 right now?
Through the Kimi app, Kimi Code, and the API (kimi-k3). Third-party hosts like Fireworks are expected once the weights ship. For what the model is and how it performs, see the Kimi K3 overview.