---
license: apache-2.0
base_model: Qwen/Qwen3.5-0.8B
library_name: mlc-llm
tags:
  - mlc-llm
  - web-llm
  - webgpu
  - game
  - npc
language:
  - en
---

# TinyAgent 0.8B (v5)

The small language model behind the AI staff in [Tiny Agents](https://tinyvoters.app/tinyagents/), a cozy pixel-art
office where Claude Code sessions show up as little coworkers. It runs entirely in the visitor's browser on WebGPU
through [WebLLM](https://github.com/mlc-ai/web-llm); nothing is sent to a server.

Each worker in the office gives it their own persona, needs, mood, work and surroundings, and it answers in short JSON:

- **Act**: pick one of a few options (lunch, coffee, the restroom, go home sick, ...) and say a short line.
- **Speak**: react to what someone just said, open a chat, or take a turn in a conversation or meeting.
- **Think**: a private one-line thought shown in the worker's info window, never said aloud.

## Details

- Fine-tune of [Qwen/Qwen3.5-0.8B](https://huggingface.co/Qwen/Qwen3.5-0.8B) (Apache-2.0), full fine-tune on about
  22,000 examples written and reviewed by Claude: decisions, reactions, conversation turns and thoughts, in the
  exact prompt format the game uses.
- MLC `q4f16_1` weights in the same layout as
  [mlc-ai/Qwen3.5-0.8B-q4f16_1-MLC](https://huggingface.co/mlc-ai/Qwen3.5-0.8B-q4f16_1-MLC), so it runs on that
  model's prebuilt WebGPU library.
- About 430 MB to download, cached by the browser after the first visit.

It's tuned for one game's prompt format and tone; it isn't a general assistant.