> For the complete documentation index, see [llms.txt](https://docs.clore.ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.clore.ai/guides/language-models/kimi-k3.md).

# Kimi K3 (2.8T — What It Takes)

What Kimi K3 is, what it actually takes to self-host 2.8 trillion parameters, and what to run on Clore.ai instead

{% hint style="danger" %}
**Straight answer first: Kimi K3 does not fit on any server listed on the Clore.ai marketplace, at any quantization.** The smallest community build is a 1-bit GGUF at **466GB**; the largest rig on the marketplace is 380GB. This page explains what K3 is, what hosting it really requires, and which models give you most of the capability on hardware you can rent today.
{% endhint %}

Moonshot AI published **Kimi K3** weights on **27 July 2026** — the first open model in the 3-trillion-parameter class. It is a genuine milestone, and it is also a useful calibration point for how far "open weights" and "self-hostable" have drifted apart.

### Key Specs

| Property     | Value                                                                                                     |
| ------------ | --------------------------------------------------------------------------------------------------------- |
| Parameters   | 2.8T total, 16 of 896 experts active per token                                                            |
| Architecture | Kimi Delta Attention (KDA) + Attention Residuals, Stable LatentMoE, 69 KDA + 24 Gated MLA layers          |
| Modality     | Text, image, video in → text out                                                                          |
| Context      | 1,000,000 tokens                                                                                          |
| License      | Kimi K3 licence (custom — commercial inference above roughly $20M annual revenue requires separate terms) |
| Release      | Hosted 16 July 2026, weights 27 July 2026                                                                 |
| Weights      | Q8 1,561GB · Q4\_K\_XL 1,509GB · Q2\_K\_XL 861GB · IQ1\_S 594GB · Q1\_0 466GB                             |

Moonshot claims roughly **2.5× better scaling efficiency than Kimi K2** from the sparser MoE and the KDA attention stack, with native multimodal training rather than a bolted-on vision tower.

## What hosting it actually takes

| Build                    | Size    | Hardware needed             |
| ------------------------ | ------- | --------------------------- |
| `UD-Q8_K_XL`             | 1,561GB | \~20× H100 80GB             |
| `UD-Q4_K_XL`             | 1,509GB | \~20× H100 80GB             |
| `UD-Q2_K_XL`             | 861GB   | \~11× H100 80GB, or 8× H200 |
| `UD-IQ1_S`               | 594GB   | \~8× H100 80GB              |
| `UD-Q1_0`                | 466GB   | \~6× H100 80GB              |
| `mgoin/Kimi-K3-pruned75` | 475GB   | \~6× H100 80GB              |

For reference, the largest Clore.ai marketplace rigs are 4× RTX PRO 6000 Blackwell (380GB) and 11× RTX 5090 (341GB). Even the 1-bit build overshoots them, and a 1-bit quantization of a frontier model is not a serious production target anyway.

{% hint style="info" %}
**If you genuinely need to serve K3**, that is a [bare metal](https://clore.ai/bare-metal) conversation — dedicated H200 or B200 nodes, provisioned on request rather than rented per minute. Check the licence first: the terms gate commercial inference above roughly $20M in annual revenue.
{% endhint %}

## What to run on Clore.ai instead

| If you wanted K3 for…      | Run this                                                                                | Fits on                      |
| -------------------------- | --------------------------------------------------------------------------------------- | ---------------------------- |
| Frontier coding and agents | [GLM-5.2](/guides/language-models/glm-5-2.md) — MIT, 1M ctx, top open coding scores     | 10× RTX 5090 rig, \~$5.00/hr |
| Long-context multimodal    | [MiniMax M3](/guides/language-models/minimax-m3.md) — 1M ctx, native video              | 8× RTX 5090, \~$2.00–3.50/hr |
| Open reasoning at scale    | [Nemotron 3 Ultra](/guides/language-models/nemotron-3-ultra.md) — 550B, OpenMDW licence | 4× RTX PRO 6000, \~$5.00/hr  |
| One model on one card      | [Qwen3.8-27B](/guides/language-models/qwen38-27b.md) — Apache 2.0, vision, 262K ctx     | 1× RTX 4090, $0.14–0.42/hr   |
| Best quality per rented GB | [Mistral Small 4](/guides/language-models/mistral-small4.md) — 119B / 6.5B active       | 2× RTX 4090, $0.28–0.84/hr   |

The honest summary: since June 2026 the open frontier has split. Models like GLM-5.2 and Nemotron 3 Ultra sit at the edge of what a big consumer rig can hold; K3 sits past it. Most of the practical gap between them closes with a good harness and a long context, both of which the models above give you.

## Using K3 without hosting it

Moonshot serves K3 through the Kimi app, Kimi Code and its API. If you are building on Clore.ai, a common pattern is to keep the expensive frontier calls on an API and run the high-volume path locally on a rented GPU — see the cost comparison in [Gemini 3.1 Flash Lite](/guides/language-models/gemini-3-1-flash-lite.md), which walks through when self-hosting actually wins.

## Next Steps

* [GLM-5.2](/guides/language-models/glm-5-2.md) — the largest model you can realistically rent here
* [GPU Pricing & Availability](/guides/getting-started/pricing.md) — what the ceiling is right now
* [Qwen3.8-27B](/guides/language-models/qwen38-27b.md) — the single-card workhorse
* [Kimi K2.5](/guides/language-models/kimi-k2.md) — the previous generation, far more tractable

### Links

* [Kimi K3 on Hugging Face](https://huggingface.co/moonshotai/Kimi-K3) · [GGUF builds](https://huggingface.co/unsloth/Kimi-K3-GGUF)
* [Kimi K3 licence](https://huggingface.co/moonshotai/Kimi-K3/blob/main/LICENSE)
* **Rent a GPU:** [Marketplace](https://clore.ai/marketplace) · [Bare metal](https://clore.ai/bare-metal)


---

# Agent Instructions
This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com.

## Querying This Documentation
If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question.

Perform an HTTP GET request on the current page URL with the `ask` query parameter, and the optional `goal` query parameter:

```
GET https://docs.clore.ai/guides/language-models/kimi-k3.md?ask=<question>&goal=<endgoal>
```

`ask` is the immediate question: it should be specific, self-contained, and written in natural language.
`goal` is optional and describes the broader end goal you are ultimately trying to accomplish on behalf of the user. GitBook uses it to tailor the answer towards what is most useful for that goal.

The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.

Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.