Brimkern
WebGPU100% on-device · Private · Zero server calls

Powerful AI models.
Directly in your browser.

Brimkern runs open-source models straight from Hugging Face on your own GPU: no installation, no server, and no subscription. Weights stream in once, stay on your device, and work completely offline.

Start chatting — Free & localExplore docs & models
curl -fsSL https://brimkern.com/install.sh | bash

macOS & LinuxCLI guideOr open in browser

brimkern-engine.ts
WebGPU v0.9
# Run any Hugging Face model directly on client GPU
$ brimkern run Qwen/Qwen2.5-0.5B-Instruct-GGUF
[gpu]  WebGPU adapter: Apple M-series (resident KV cache)
[http] Streaming Q4_K_M weights (378 MB) via range requests
[wgsl] 14 custom shaders compiled · Zero server calls
✓ Model ready in 1.4s · 47.2 tok/s decode
Launch Qwen 2.5 in chat
01 Instant trial · 1 click

Pick a model. Chat immediately.

⚡ Fastest149 MB
LFM2.5 230M

Liquid AI architecture. Instant download, minimal memory footprint.

Arch: Liquid NetWeights: .brikSpeed: Instant
Launch in chat
🧠 Fluent378 MB
Qwen 2.5 0.5B

Alibaba. Strong at reasoning, coding and multilingual chat.

Arch: TransformerFormat: Q4_K_MCtx: 32,768
Launch in chat
📦 Compact270 MB
Gemma 3 270M

Google Gemma 3. Highly capable compact model for general queries.

Arch: Gemma 3Format: Q4_K_SCtx: 8,192
Launch in chat
02 Architecture & Privacy

Three reasons to run AI directly in your browser

privacy first

100% Private & offline

Your conversations and documents never leave your machine. No accounts, no servers, and no telemetry. Once loaded, models continue to work even without an internet connection.

instant streaming

Zero install, zero configuration

No Python, Docker, or complex drivers to configure. Models stream in seconds through standard HTTP ranges and are cached locally on your device for immediate future access.

open ecosystem

Any open-source model

Run standard GGUF and BRIK models from Hugging Face: Qwen, Gemma, Llama and more. Generate text, reason through problems, describe photos, or create images.

How this compares to WebLLM (measured benchmarks)

03 Pipeline & Shaders

From model weights to answers on your GPU

  1. 01
    You select author/model

    A one-click curated preset or any Hugging Face repo ID or direct link.

  2. 02
    We resolve the best weights

    The Hub API lists the repo; the best quantization is selected automatically.

  3. 03
    It streams into local cache

    Weights stream progressively in the background and stay cached on your device.

  4. 04
    It runs on your GPU

    WebGPU compute shaders execute the model live. Zero data leaves your machine.

04 Hardware Benchmarks

Measured on real hardware

149 MB
lightest chat model, cached once
47.2 tok/s
prefill on a 7B int4 (WebLLM: 18.7)
15.8 s
to reload 4.7 GB from local cache
0
servers, accounts, or API keys needed
05 for your own product

One script tag, an assistant that costs nothing to run

The compute is your visitor’s GPU: no inference bill, no rate limit, no data leaving their browser. The model only downloads when someone actually opens the widget, so your page speed is untouched.

SDK page & live demo 
<script src="https://brimkern.com/sdk.js"></script>
<script>
  Brimkern.embed({
    system: "You answer questions about my shop.",
  });
</script>
06 also in your terminal

The same engine, as a coding assistant for your shell

It reads the project you are in and runs Qwen 3 4B on your GPU. Nothing leaves the machine.

See the CLI
07 also in the box

One engine, four modalities

  • Chat: multi-turn, reasoning models, French & English, on a resident GPU KV cache.
  • Vision: attach an image and ask about it (Qwen2-VL, desktop).
  • Images: text-to-image in the tab (SD-Turbo / SDXS with a WebGPU diffusion stack).
  • Video (beta): short animated clips from a prompt, on the same kernels.