# InternVL3_5-GPT-OSS-20B-A4B-Preview-HF GPU requirements

InternVL3_5-GPT-OSS-20B-A4B-Preview-HF has 21B parameters. To serve it, plan for about 51 GB of GPU memory in 16-bit, 25 GB in 8-bit or 13 GB in 4-bit, including about 20% for the KV cache and activations.

It is a mixture of experts: 32 experts, 4 of them active for each token. All of them stay in GPU memory, so plan memory for all 21B parameters.

[Deploy it on Arrakis](/console/#/models?model=OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview-HF).

## What it takes to serve it

| Precision | Memory and the smallest setup |
|---|---|
| 16-bit | 51 GB: one [H100](/gpus/h100.html.md) or [A100](/gpus/a100.html.md), H100 from $2.25 per GPU-hour |
| 8-bit | 25 GB: one [RTX 5090](/gpus/rtx-5090.html.md), from $0.58 per GPU-hour |
| 4-bit | 13 GB: one [RTX 4090](/gpus/rtx-4090.html.md) or [RTX 3090](/gpus/rtx-3090.html.md), RTX 3090 from $0.19 per GPU-hour |

## What it takes to fine-tune it

| Method | Memory and the smallest setup |
|---|---|
| QLoRA | 15 GB: one [RTX 4090](/gpus/rtx-4090.html.md) or [RTX 3090](/gpus/rtx-3090.html.md), RTX 3090 from $0.19 per GPU-hour |
| LoRA | 51 GB: one [H100](/gpus/h100.html.md) or [A100](/gpus/a100.html.md), H100 from $2.25 per GPU-hour |
| Full fine-tuning | 340 GB: 2 × [B200](/gpus/b200.html.md), from $13.58 per hour for all 2 |

## About the model

| | |
|---|---|
| Parameters | `21B` |
| Active for each token | `4B` |
| Experts | `32` |
| Experts for each token | `4` |
| Context window (tokens) | `131,072` |
| Published in | `BF16` |
| License | `apache-2.0` |
| Released | `2025-08-29` |
| Weights | [OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview-HF](https://huggingface.co/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview-HF) |

## Serve it yourself

```
vllm serve OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview-HF --tensor-parallel-size 1
```

## Questions

**How much GPU memory does InternVL3_5-GPT-OSS-20B-A4B-Preview-HF need?** To serve it, about 51 GB in 16-bit, 25 GB in 8-bit, 13 GB in 4-bit, including about 20% for the KV cache and activations. To fine-tune it, about 15 GB with QLoRA, 51 GB with LoRA and 340 GB for full fine-tuning with Adam, before activations.

**Can I run InternVL3_5-GPT-OSS-20B-A4B-Preview-HF on one GPU?** Yes. In 16-bit it needs about 51 GB, which fits on one H100 or A100.

**What does it cost to run InternVL3_5-GPT-OSS-20B-A4B-Preview-HF?** In 16-bit it needs one H100 or A100; in 4-bit, one RTX 4090 or RTX 3090. Each setup on this page shows the lowest price per GPU-hour on Arrakis right now, and you pay by the hour from a prepaid balance.

**How do I deploy InternVL3_5-GPT-OSS-20B-A4B-Preview-HF on Arrakis?** Open Models in the Arrakis console, paste OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview-HF, choose vLLM, SGLang or llama.cpp, and rent a GPU it fits on. You get an OpenAI-compatible endpoint with its own API key, and root SSH to the machine.


More models: [/models](/models.html.md).