Forem: Glows.ai

How to Download HuggingFace Models on Glows.ai

Glows.ai — Thu, 02 Apr 2026 15:25:57 +0000

This tutorial explains how to download HuggingFace models on Glows.ai, using two available methods: Glows.ai Datadrive storage (local download and upload), and instance-based storage (download directly inside an instance).

Glows.ai Datadrive storage: Data can be read and written without limits as long as your Storage Space Plan remains valid. Download speed depends on your local network, making this option suitable for users who frequently need to download the same data (e.g., for model serving).
Instance-based storage: Data is valid only while the instance is running; once the instance is released, the data will be deleted. Because it shares the bandwidth of the instance’s data center, download speeds are fast. This method is suitable for users who need the data for one-time use (e.g., testing model performance).

Glows.ai Datadrive Storage

This method saves data to Glows.ai Datadrive. Make sure your Space Storage plan provides enough capacity, and allocate that capacity to the DataDrive in the region you intend to use.

Allocate Storage

Suppose you need to download a model of 65GB and plan to use an NVIDIA GeForce RTX 4090 GPU in the TW-03 region. You’ll first need to go to Storage Space and purchase a 100GB storage package.

Then, click the Modify button in the Storage Space interface to allocate 70GB of space to the TW-03 region Datadrive.

Datadrive Client Download

The Data Drive client currently supports downloading models directly from HuggingFace to the corresponding Datadrive of different region.
The process works as follows: using your local network, the client downloads HuggingFace model chunks locally, then synchronizes them to the Datadrive.

Install the Data Drive client: Download here
Follow the tutorial: Download models from HuggingFace

Instance-based Storage

Create an Instance

This method requires creating an instance on Glows.ai. Suppose you would use an NVIDIA GeForce RTX 4090 GPU in the TW-03 region, with the environment CUDA 12.8 Torch 2.8.0 Base.

Once the instance is created, you can connect to it via SSH or HTTP Port 8888 (JupyterLab).

Download Model Using Commands

JupyterLab is simple to use. The following example demonstrates operations within JupyterLab.
Open a new Terminal.

Enter the following command to install HuggingFace’s official model management tool huggingface_hub:

pip install -U huggingface_hub

Once installed, you can use the hf command to download model files directly to the instance.
For example, to download openai/gpt-oss-20b into the /gpt-oss-20b directory, use:

hf download openai/gpt-oss-20b --local-dir /gpt-oss-20b

Running HuggingFace Models on Glows.ai

Some frameworks support directly loading and running HuggingFace models, such as Transformers, SGLang, and GPUStack.
You can use the software you’re most familiar with for deployment or refer to the tutorials below:

HuggingFace official website also provides usage examples.
If you have any questions or suggestions during your implementation on Glows.ai, feel free to contact us.

How to Run OpenClaw on Glows.ai

Glows.ai — Mon, 23 Mar 2026 09:30:55 +0000

This tutorial will walk you through how to rent an NVIDIA GeForce RTX 4090 GPU on Glows.ai to run OpenClaw with a local model, and share a safer and more convenient deployment approach.

This article includes the following content:

How to create an instance on Glows.ai
How to configure a local model with OpenClaw
How to connect to and use OpenClaw

OpenClaw is an open-source AI Agent framework that has gained rapid popularity recently. It is designed to combine large language models with practical tool capabilities, enabling AI not only to answer questions but also to execute real tasks on computers or servers. It typically runs in a self-hosted manner and can be deployed on local devices or in cloud environments, while interacting with users through APIs or chat applications.

Unlike traditional conversational AI, OpenClaw functions more like an intelligent assistant with action capabilities: the AI model is responsible for understanding goals and making decisions, while OpenClaw handles tool orchestration, command execution, and overall workflow management.

Now let’s dive into a hands-on example together.

Create an Instance

We create an instance on demand on Glows.ai. You can refer to the tutorial. Please make sure to use the officially preconfigured OpenClaw(img-6np58jp2) image.

On the Create New page, select Inference GPU -- 4090 for Workload Type, then choose the OpenClaw image first. This image has already been preconfigured with the required environment.

Datadrive is a cloud storage service provided by Glows.ai for users. Before creating an instance, users can upload the data, models, code, and other content they want to run to Datadrive. When creating the instance, simply click the Mount button in the interface to mount Datadrive to the instance being created. This allows us to directly read and write Datadrive content from within the instance.

In this tutorial, we only run inference services, so mounting Datadrive is not required.

Once everything is ready, click Complete Checkout in the lower-right corner to complete instance creation.

The estimated startup time for an OpenClaw image instance is 30–60 seconds. We can view the instance status and related information in the My Instances interface. After the instance starts successfully, we will see the following information:

SSH Port 22 is the SSH connection for the instance
HTTP Port 8888 is the JupyterLab connection
HTTP Port 11434 is the Ollama API connection

Connect to the Instance

Visit the link corresponding to HTTP Port 8888 on the instance page to open the JupyterLab service. Then follow the illustration to create a new Terminal.

Upgrade OpenClaw

Thanks to the open-source community, OpenClaw is iterating very quickly. You can upgrade to the latest version to obtain the full feature set.

First, enter the following commands in the Terminal to check the version status. You will see the latest available version number.

openclaw --version
openclaw update status

Then enter the following command to update the version.

openclaw update

Reminder: The final systemctl command error can be ignored. This is because systemctl is not supported inside Docker containers. The tutorial below will explain how to start the OpenClaw service manually.

Basic OpenClaw Configuration

First, enter the following command to open the OpenClaw configuration interface, mainly for agreeing to the OpenClaw terms and configuring the Telegram API Token.

openclaw onboard

Use the default options. The model configuration can be left empty for now, as we will configure a local Ollama model later. You may also configure a third-party model API as needed.

You can configure the communication software according to your needs. This tutorial uses Telegram as an example. First, you need to send the following command to @BotFather on Telegram to create a new Bot and obtain the Bot API Token.

/newbot

Then enter the Telegram Bot API Token in the OpenClaw configuration interface.

Finally, when you see Onboarding complete, the configuration is complete.

OpenClaw Local Model Configuration

With Ollama, you only need a single command to download and configure a model for OpenClaw. The following command uses the qwen3.5:4b model as an example.

ollama launch openclaw --model qwen3.5:4b

After downloading the model, the OpenClaw service will start automatically. We can then begin using OpenClaw.

Reminder: Remember the link under Open the Web UI here, especially the token. It will be needed later when we access the Web interface.

Three Entry Points for Interacting with OpenClaw

Chat in the CLI

After the previous command finishes running, you can directly chat with OpenClaw in the interface. For example, here we ask it to create a Python file and write the Fibonacci algorithm into it.

Chat in Telegram

First, send a message to the Telegram Bot created earlier. After a short while, you will receive an authentication command.

Copy the authentication command above and execute it in the instance Terminal. After execution is complete, you can formally use OpenClaw in Telegram.

To test its memory capability, we can directly ask it for the path of the Fibonacci algorithm file we asked it to write earlier. As shown in the screenshot, the Bot quickly replies with both the path and the source code content.

Chat in a Web Browser

For security reasons, OpenClaw binds its service host to 127.0.0.1 by default. We need to use SSH port forwarding to forward the OpenClaw service port from the instance to a local computer port before accessing it.

On the My Instance page in Glows.ai, you can view the SSH information. We need to convert the SSH Command into an SSH port forwarding command.

The modified command is as follows. You only need to replace these parts of the command.

Replace 2xxx7 with the corresponding value in the SSH Command
Replace 18888 with any available port number on your local computer
Replace root@tw-05.access.glows.ai with the corresponding value in the SSH Command

ssh -p 2xxx7 -NL 18888:localhost:18789 root@tw-05.access.glows.ai

Run the modified SSH port forwarding command in your local Terminal/CMD. After pressing Enter, paste the SSH Password. It is normal that nothing is displayed after pasting the password. Press Enter again. If no errors appear, the message has been forwarded successfully.

Reminder: If you are a Windows user, it is recommended to download and install Windows Git locally.

After the forwarding succeeds, visit http://localhost:18888/#token=xxxxxxxx locally. Please replace the token value with the token displayed when starting the OpenClaw service in OpenClaw Local Model Configuration.