---
title: Supported models
description: Model families vLLM supports natively, and how to check whether a specific model works.
url: https://vllm-agent-docs.sudhanvasp.dev/reference/supported-models
---

# Supported models

Model families vLLM supports natively, and how to check whether a specific model works.

> **Note:**
  Adapted from vLLM's [Supported Models reference](https://docs.vllm.ai/en/latest/models/supported_models/) (Apache-2.0).

vLLM natively implements over 100 model architectures, and supports hundreds
more through the Transformers modeling backend as a fallback. Models fall
into three categories:

- **Text-only generative models** — standard causal language models
- **Multimodal generative models** — text combined with image, video, and/or audio input
- **Pooling models** — embedding, reranking, and classification models

## Popular supported families

- Llama (Llama 2, Llama 3.x)
- Qwen (Qwen2, Qwen3, QwQ)
- Mistral / Mixtral
- DeepSeek (V2, V3, and newer)
- Gemma
- Phi
- Falcon
- InternLM
- ChatGLM
- Jamba
- Baichuan
- OPT

This list is representative, not exhaustive — check a specific model with
one of the methods below rather than assuming absence from this list means
it's unsupported.

## Check whether a model is supported

**Fastest: read the model's `config.json`.** Open the model's page on
Hugging Face Hub and check the `"architectures"` field. If it matches an
architecture vLLM implements, it's natively supported.

**Authoritative: try loading it.**

```python
from vllm import LLM
llm = LLM(model="your-org/your-model", runner="generate")
output = llm.generate("Hello, my name is")
```

If the architecture isn't recognized, vLLM raises
`Model architectures [...] are not supported` at load time — see
[Troubleshooting](/troubleshooting#model-architectures-are-not-supported)
if you hit this on a model you believe should work.

## Custom or gated models

Models that ship custom modeling code on the Hub need
`--trust-remote-code` passed to [`vllm serve`](/reference/server-args).
Gated models need a Hugging Face token with access, set via `HF_TOKEN`.