# Model Serving - Sandbox Models
This is the list of AI models we are currently **testing and evaluating**. Models listed here are not yet generally available, but they are being validated for performance, stability, and cost. Once a model passes our evaluation, it will be promoted to the [Active Models](active-models.md) list.
These are models we are **looking into and considering** for the production catalog. They are available on a **trial basis** with **no SLA**. Pricing and throughput may not be listed yet for models that have not been fully tested; where figures are shown, they are preliminary and may change before promotion to [Active Models](active-models.md).
For explanations of recent additions and removals, see the [Changelog](../changelog/index.md).
---
## Sandbox Models
| Model | Type | Description |
|---------------------|-----------|----------------------------------------------|
| [zai-org/GLM-5.3](https://huggingface.co/zai-org/GLM-5.3) | Reasoning | Open-weight successor to GLM 5.2 (753B parameters; same base model, post-training gains). Optimized for complex coding and long-horizon agentic tasks, with stronger software-engineering and terminal performance than GLM 5.2. |
| [moonshotai/Kimi-K3](https://huggingface.co/moonshotai/Kimi-K3) | Multimodal | Open-weight multimodal agentic model (2.8T parameters, 104B activated) optimized for long-horizon coding, agentic knowledge work, and reasoning. Native text, image, and video support with a 1-million-token context window. |
| [MiniMaxAI/MiniMax-H3](https://huggingface.co/MiniMaxAI/MiniMax-H3) | Multimodal Video generation | Omni-modal generative system for unified text, image, video, and audio understanding. Generates video with native stereo audio at up to 2K resolution and 4–15 second durations. Supports text-to-video, first/last-frame-to-video, and reference-to-video workflows. |