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

# Open-Weight Models

> Learn which AI Gateway models are open-weight and what that means for how you can use them.

An open-weight model is one whose author publishes the trained weights for anyone to download.
You can call it through a hosted provider, like any other model, or run the same weights on your own hardware.

## What open-weight means

The weights are the parameters a model learns during training.
With a closed model, such as GPT-5.5 or Claude, only the author and its partners can run the weights, so you can only reach the model through their APIs.
With an open-weight model, the author releases the weights under a license, and any provider or person can serve them.

Open-weight isn't the same as open source.
Authors often publish the weights without the training data or training code.
Licenses also differ: some, like Apache 2.0 and MIT, allow almost any use, while others, like the Llama Community License, add conditions.
Check the license on the author's model card before you use a model commercially or fine-tune it.

## What it means in the AI Gateway

You send requests to open-weight models the same way you send them to closed models.
Being open-weight changes where the model can run:

* **More than one provider can serve it.** When several providers in the [model catalog](https://ngrok.com/docs/ai-gateway/reference/model-catalog.md) serve the same model, the gateway can [fail over](https://ngrok.com/docs/ai-gateway/guides/configure-fallback-models.md) between them.
  Use a `provider:model` prefix to keep requests on one provider.
* **You can run it yourself.** Serve the weights with a tool like Ollama or vLLM, then [add it as a custom provider](https://ngrok.com/docs/ai-gateway/guides/use-a-model-you-run-yourself.md).
  Your apps call it through the same gateway endpoint and access keys as your hosted models.
* **You can mix self-hosted and hosted copies.** List your own server first and a hosted provider as a fallback, so requests keep working when your server is down.

For example, this request tries Qwen3-32B on your own Ollama server first, then on Groq:

```json theme={null}
{
  "model": "ollama:qwen3:32b",
  "models": ["groq:qwen/qwen3-32b"],
  "messages": [{"role": "user", "content": "Hello"}]
}
```

## Open-weight models in the catalog

These models in the [model catalog](https://ngrok.com/docs/ai-gateway/reference/model-catalog.md) are open-weight.
Faster or provider-only variants, such as `-highspeed`, `-x`, `-flashx`, and `-turbo` models, and the Qwen Flash, Plus, and Max tiers, don't have published weights.

| Author | Model IDs | Providers |
| - | - | - |
| OpenAI | `openai/gpt-oss-120b`, `openai/gpt-oss-20b`, `openai/gpt-oss-safeguard-20b` | Groq |
| Meta | `meta-llama/llama-4-scout-17b-16e-instruct`, `meta-llama/llama-prompt-guard-2-86m`, `meta-llama/llama-prompt-guard-2-22m` | Groq |
| DeepSeek | `deepseek/deepseek-v4-pro` | DeepSeek, QwenCloud, Tencent TokenHub |
| DeepSeek | `deepseek/deepseek-v4-flash` | DeepSeek, Tencent TokenHub |
| DeepSeek | `deepseek/deepseek-v4-flash-0731` | Tencent TokenHub |
| DeepSeek | `deepseek/deepseek-v4.1-flash` | QwenCloud |
| Alibaba Qwen | `qwen/qwen3-32b` | Groq |
| Alibaba Qwen | `qwen/qwen3.8-2.4t-a95b` | QwenCloud |
| Moonshot AI | `moonshotai/kimi-k3` | Moonshot AI, QwenCloud, Tencent TokenHub |
| Moonshot AI | `moonshotai/kimi-k2.7-code`, `moonshotai/kimi-k2.6` | Moonshot AI, Tencent TokenHub |
| Moonshot AI | `moonshotai/kimi-k3.1-pro` | Moonshot AI |
| Z.ai | `z-ai/glm-5.3` | Z.ai, Mistral AI, QwenCloud, Tencent TokenHub |
| Z.ai | `z-ai/glm-5.2` | Z.ai, QwenCloud, Tencent TokenHub |
| Z.ai | `z-ai/glm-5.3-flash`, `z-ai/glm-5.1`, `z-ai/glm-5` | Z.ai, Tencent TokenHub |
| Z.ai | `z-ai/glm-4.7`, `z-ai/glm-4.7-flash`, `z-ai/glm-4.6`, `z-ai/glm-4.5`, `z-ai/glm-4.5-air`, `z-ai/glm-image` | Z.ai |
| MiniMax | `minimax/minimax-m3`, `minimax/minimax-m2.7` | MiniMax, Tencent TokenHub |
| MiniMax | `minimax/minimax-m3.1-flash`, `minimax/minimax-m2.5`, `minimax/minimax-m2.1`, `minimax/minimax-m2` | MiniMax |
| Xiaomi | `xiaomi/mimo-v2.6-pro`, `xiaomi/mimo-v2.6-flash`, `xiaomi/mimo-v2.5-pro` | Tencent TokenHub |
| Tencent | `tencent/hy4-preview`, `tencent/hy3`, `tencent/hy-mt2-pro`, `tencent/hy-mt2-plus`, `tencent/hy-mt2-lite` | Tencent TokenHub |
| StepFun | `stepfun/step-5-preview` | Tencent TokenHub |

## Next steps

* [Models](https://ngrok.com/docs/ai-gateway/concepts/models.md): How the gateway names and resolves models
* [Use a model you run yourself](https://ngrok.com/docs/ai-gateway/guides/use-a-model-you-run-yourself.md): Connect Ollama, vLLM, or LM Studio
* [Configure fallback models](https://ngrok.com/docs/ai-gateway/guides/configure-fallback-models.md): Try another model when the first one fails


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.