OB Studio

Try Tot-360M live

Every answer here comes from a Raspberry Pi 4 (4 GB) on my desk, not a cloud GPU. Pick a checkpoint on the left. The Pi keeps one model in memory and can answer two people at the same time. If someone else is using a different model, you wait in line until they finish and the Pi swaps.

Connecting to the Pi...One model in RAM at a time · up to 2 people at once
Model
    Effort

    Up to 2,048 tokens · balanced chat & games · temp 0.2

    Q4_K_M · ~4.85 bits/weight241 MiB (63.5% quantized from 660 MiB FP16)

    This is a raw base checkpoint. It does not chat. Type the start of some text and it keeps writing.

    What to expect

    This is a 345M-parameter model. It is good at short code snippets, simple HTML5 games, and bilingual small talk. It makes things up, gets facts wrong, and loses track in long conversations. SFT v1 is the first checkpoint trained to follow chat turns; the others are base models.

    How it runs

    Stock llama-server with two parallel slots of 4,096 tokens each. Every checkpoint runs in Q4_K_M (about 4.85 bits per weight, 241 MiB, 63.5% quantized from the 660 MiB FP16 weights). Effort adjusts token budget and sampling temperature from fast concise replies (Low, 1,024 tokens) to complete single-file HTML5 code (Medium 2,048 / High 2,560 tokens). The model unloads after 10 minutes with no traffic.

    Run it yourself

    The same files are free to download and run offline on a phone, laptop, or your own Pi. Downloads and training logs.