╔══════════════════════════════════════════╗
 ║   ACCESS CONTROL — local-llm-playground  ║
 ╚══════════════════════════════════════════╝
$ authenticate --mode=hackathon
~/local-llm-playground $
airgap_ok │ net: OFF │ gpu: WebGPU │ │

cat ./README.md

> on-prem AI — no internet, no admin rights, no data leaves your machine.

Seven self-contained examples that run entirely in your browser (via WebGPU) or against a local Ollama server. Start with a simple chat and work your way up to OCR-powered structured extraction, multi-model comparison, and prompt engineering.

07 examples│ 0 deps│ 0 tracking│ ~270M–7B params│ license: MIT
getting_started [2]
webgpu_path examples 01–05, 07 // recommended
  1. Use Chrome 113+ or Edge 113+ — pre-installed on most machines, no install needed.
  2. Download the .task model files — see 01_chat for Kaggle links. Both files are also pre-loaded on the hackathon machines.
  3. Open the HTML file and pick the model when prompted. That’s it.
ollama_path examples 06, 07 // optional
  1. Download Ollama — single binary, runs without admin on macOS / Linux.
  2. Run ollama pull gemma3:1b in a terminal to fetch a model.
  3. Open 06_ollama.html or 07_prompt.html — connects to localhost:11434 automatically.
models [2]
gemma3_270m fast // q4_0-web.task
params 270 M
quant 4-bit (q4_0) — smaller file, some quality loss
file size~500 MB
context 32K tokens
speed 20–35 tok/s on a modern laptop GPU
used in 01, 02, 03, 04, 05, 07
Very fast — responses feel near-instant Good at classification and structured extraction Struggles with multi-step reasoning and code Inconsistent on open-ended generation
gemma3_1b smarter // q8_ekv4096.task
params 1 B — ~3.7× more capacity than 270M
quant 8-bit (q8) — near-lossless quality
file size~1 GB
context 4096-token KV cache (sliding window)
speed 8–15 tok/s on a modern laptop GPU
used in 04, 07
Noticeably better reasoning and instruction-following q8 preserves near-full model quality Slower — pairs well with the 270M for comparison (04) Still limited on complex coding or math tasks
examples [7]
01

basic_chat_

Type a message and get a response from Gemma 3 270M running entirely in your browser.

beginner WebGPU
→
02

text_classifier

Define your own categories, paste any text, and let the model assign it to the right bucket.

beginner WebGPU
→
03

data_extractor

Paste unstructured text (receipts, emails, contacts) and get clean structured JSON out.

intermediate WebGPU
→
04

model_compare

Run the same prompt on Gemma 3 270M and 1B side by side and compare quality and speed.

intermediate WebGPU
→
05

ocr_extract

Drop in a screenshot of a form, invoice or ID card — OCR + LLM returns typed fields.

advanced WebGPU
→
06

multi_model_compare

Send the same prompt to several local Ollama models in parallel and diff outputs side‑by‑side.

advanced Ollama
→
07

prompt_lab

Same message, three system prompts side by side. See how framing changes everything — the fastest skill to learn in AI.

intermediate Ollama WebGPU
→
secrets [?]
hidden_commands // click to reveal ▶

// faggruppemøte_mai_2026

Open Weights LLM — Knowledge Check
10 questions · 3 options each · instant feedback
🔒
Quiz locked
Visit all 7 example sessions to unlock the knowledge check.
01✓
02✓
03✓
04✓
05✓
06✓
07✓
0 / 7 visited