sts TeamGive piko a Brain
Five model tiers, installation, and sampling tuning
On this page
Give piko a Brain: the guide
By default piko answers from a built-in quote book; install a local brain and it can actually converse. The model runs inference in your browser and nothing you say is uploaded. Entry point: /piko.
Pick a tier
| Tier | Size | Best for |
|---|---|---|
| piko-0.5B (the original) | 271MB | The author's own fine-tune: piko's personality and a map of the whole site are baked in. The default pick |
| Qwen2.5-0.5B | 469MB | A general-purpose base guest-starring; small and quick |
| Qwen2.5-Coder-0.5B | 469MB | The piko that writes code |
| Qwen2.5-1.5B | 882MB | Smarter but slower; better on desktop |
| Qwen2.5-3B | 2007MB | The desktop flagship; needs 2GB+ of free memory. Not for phones |
The difference between the original and the guests: the original brain was fine-tuned on the site's own Q&A material, so it talks like piko out of the box; the other four are general models acting the part through prompts. You can install all five - switching between cached tiers is instant. Use the original for casual chat, and switch to the Coder tier when you want it to read code.
Installation
- Open the "Brain" button next to the input box
- Pick a tier and confirm the download (the model caches in your browser; you only download once)
- Wait for the download and load to finish - piko then switches into conversation mode automatically
Download sources
The general Qwen tiers come from two mirrors, ModelScope and hf-mirror, and a failed download automatically retries on the other one. The original tier's repository lives on ModelScope only; failed downloads resume from where they stopped. File contents match the published repos.
Usage and tuning
- Once installed, every visit to /piko wakes the brain automatically - no re-download
- The original tier's sampling defaults follow its model card (temperature 0.7, repetition penalty 1.05). The defaults in "Sampling" track the current tier, so untouched settings are safe to use
- Answers too random, or repeating themselves? Open "Sampling" in the brain panel: temperature, min_p, top_k, and repetition penalty are all adjustable. For 0.5B, lower the temperature a bit; min_p and the repetition penalty cut long-tail tokens and stop loops
- To change tiers just pick another one in the panel - cached tiers switch instantly
Tuning intuition: high temperature gets playful, low temperature gets dutiful; if the answer goes in circles, nudge the repetition penalty up. Change one parameter at a time - that's the fastest way to feel the difference.
Clearing the cache
"Clear brain cache" in the top right, confirm twice, and it deletes every downloaded model and engine file to free disk space. The page refreshes itself afterwards.
FAQ
The phone says it "can't run the brain", or a WebGPU error appears
Some mobile browsers have older engines, and the brain's startup stalls while probing GPU info. When you see that, chat in quote mode for now; to install a brain, update your browser or open /piko in Chrome or Edge and try again.
Privacy
Model files come from public hosting platforms (ModelScope and hf-mirror) and cache in your browser's local storage. Inference happens entirely on your device - no data is uploaded.