MiniCPM5 WebGPU

MiniCPM5 WebGPU

Runs the MiniCPM5 2B model on your own GPU in a browser tab, to chat and to make websites, Word documents, Excel workbooks and PDFs. For people who want a local assistant with nothing to install.

Made by ProCreations on Hugging Face. These notes are Mindket's, written from its Space page and README; the app's own files and license are on Hugging Face.

About

What it is: This Space runs the MiniCPM5 2B language model on your own graphics card, inside a browser tab. You chat with it and ask it to make websites, Word documents, Excel workbooks and PDFs. It uses a purpose-built WebGPU runtime, so the model work stays on your GPU.

Who it suits: It suits people who want a local chat and document assistant without installing software, and developers who want to see a model run on WebGPU.

Industry | Who uses it | How they use it | Example
Web development | Front-end developers | Study a model running on WebGPU | Look at the runtime and its benchmark notes
Office work | Anyone who drafts documents | Ask the model for a document or sheet | Request a draft workbook or a short PDF
Web design | Hobby builders | Ask for a small website | Request a one-page site and review the code
Research | Model researchers | Test speculative decoding in a browser | Compare the Auto, On and Off draft settings

You can skip it if your device has no WebGPU support or you cannot spare a multi-gigabyte download.

Where to use it:

• A web browser with WebGPU, on the Space's page at Hugging Face
• Your own computer, on Windows, macOS or Linux, by opening the downloaded files in a browser

Use case | Who | What they get back
Local chat | Anyone | Answers from a model that runs on your GPU
Document making | Office users | Websites, Word documents, Excel workbooks and PDFs
Runtime testing | Developers | A look at a WebGPU runtime and its settings

How to set it up:

• A browser that supports WebGPU and a GPU to run it
• Room for the model download, which the page puts at 3.33 GB with DSpark
• Python, only if you want to download the files with the Hugging Face tool

1. Open the Space's page on Hugging Face to use it online.
2. To keep a copy, install huggingface_hub and download the Space into a folder.
3. Move into that folder.
4. In a browser, open the index.html file inside the dist folder.

pip install -U huggingface_hub
hf download ProCreations/minicpm5-2b-webgpu --repo-type space --local-dir minicpm5-2b-webgpu
cd minicpm5-2b-webgpu

How to use it: Open the page, pick your settings on the launch screen, and start chatting. You can change the same settings later in model settings. The DSpark draft model has three modes.

Tool | What it does
Auto | Measures draft cost and switches to ordinary decoding when drafting is slower
On | Forces drafting with the DSpark model
Off | Skips downloading and allocating DSpark

Example: Ask it, "Make a one-page website for a small bakery." The model writes the page code in the chat. The README notes that one generated page needed a revision, so check what you get before you use it.

For newcomers and experts: If you are new, open the Space in a recent desktop browser and let the model download once. Then try a short chat before asking for a document.

If you know the field, the runtime is JavaScript and WGSL for a 42-layer Llama architecture, with every layer on the GPU and no WASM inference binary. It uses Q8_0 weights, and the KV cache can be set to one of four precisions, from FP32 down to 4-bit. DSpark adds speculative decoding, where a draft model proposes up to seven tokens and the main model checks them in one pass. The author says it does not always give a speedup.

The browser fetches the model files
The WebGPU runtime loads them onto your GPU
DSpark may draft several tokens ahead
The main model checks them and streams text

Good to know:

• Speed: on one Apple M4 Max in Chromium, the dedicated runtime reached 107 tok/s on prose. The author says speed varies with browser, hardware and prompt.
• License: apache-2
• Cost to run: none named. It needs a capable GPU and the model download.
• Data it reaches: the model files fetched from Hugging Face. The README says inference stays on your GPU.
• Not for: long-context quality claims. The author checked that the 64K and 128K setups allocate and generate briefly, and did not run a full quality test.

Title: Decode speed on Apple M4 Max, 256 tokens
Prose, old runtime: 58 tok/s
Prose, dedicated runtime: 107 tok/s
Prose, with DSpark Auto: 104 tok/s
JavaScript, dedicated runtime: 104 tok/s
JavaScript, with DSpark Auto: 164 tok/s

On Hugging Face, 13 November 2026

• Likes: 9
• Built with: static
• License: apache-2.0

Links

• Space: https://huggingface.co/spaces/ProCreations/minicpm5-2b-webgpu
• Files: https://huggingface.co/spaces/ProCreations/minicpm5-2b-webgpu/tree/main

Try it

Open the Space in your browser, or copy it to your own Hugging Face account from its page. Nothing in this kit runs by itself.