> ## Documentation Index
> Fetch the complete documentation index at: https://docs.tess.im/llms.txt
> Use this file to discover all available pages before exploring further.

# Kimi K3

Kimi K3 (Moonshot AI) is the largest **open-weight** frontier model in Tess: **2.8T** total parameters (**104B** active per token), **native vision**, and **1M** context. Use it for long-horizon coding, agents, and multimodal knowledge work.

| **Model IDs**<br /><br />`kimi-k3-low` · `kimi-k3-high` · `kimi-k3-max`                      | **Context**<br /><br />1M   | **Provider**<br /><br />Moonshot AI | **API launch**<br /><br />16 Jul 2026              |
| :------------------------------------------------------------------------------------------- | :-------------------------- | :---------------------------------- | :------------------------------------------------- |
| **Capabilities**<br /><br /><Icon icon="brain" /><Icon icon="image" /><Icon icon="wrench" /> | **Speed**<br /><br />Medium | **Cost**<br /><br />Medium–High     | **Intelligence**<br /><br />Frontier (open-weight) |

## What changed vs Kimi K2.6

|                           | Kimi K2.6            | Kimi K3                     |
| ------------------------- | -------------------- | --------------------------- |
| Total / active params     | —                    | **2.8T / 104B**             |
| Context                   | 1M                   | 1M                          |
| Vision                    | Yes                  | Native text + image + video |
| Reasoning in Tess         | —                    | low / high / **max**        |
| Tess credits / 100 tokens | 0.042 in / 0.196 out | **0.1596 in / 0.798 out**   |

MoE architecture (896 experts, 16 active per token). Tess exposes three reasoning levels; **Max** is the deepest (Moonshot’s default at launch).

## Pricing (Tess credits)

Values follow [Models and Costs](/en/models-and-cost) (credits per 100 tokens). Same rate across reasoning levels:

| Model        | Input / 100 tokens | Output / 100 tokens |
| ------------ | ------------------ | ------------------- |
| Kimi K3 Low  | 0.1596             | 0.798               |
| Kimi K3 High | 0.1596             | 0.798               |
| Kimi K3 Max  | 0.1596             | 0.798               |

Cached context reads are billed at a lower rate than a full input pass.

> **Screenshot placeholder — model picker:** Capture the chat model selector with **Kimi K3** (Low / High / Max) visible.

## Ideal use cases in Tess

1. Long-horizon coding — large repos, multi-file refactors, code agents
2. Persistent agent workflows with tool loops
3. Multimodal knowledge work (documents, images, and video in one context)
4. Deep reasoning with **Kimi K3 Max**
5. Open-weight frontier workloads that still need frontier-class capacity
6. Long sessions (1M context) without switching models

<Tip>
  **Best practices**

  * Use **Low** for drafting and high-volume passes; **High** as the default; reserve **Max** for planning-heavy tasks.
  * K3 costs more than K2.6 — don’t swap it in as a blanket default on cheap extraction jobs.
  * Lean on cached context in long threads to keep input cost down.
</Tip>

See also: [Models and Costs](/en/models-and-cost) · [Kimi K3 (Moonshot)](https://www.kimi.com/blog/kimi-k3).