---
title: Thousand Token Wood
emoji: 🍄
colorFrom: green
colorTo: yellow
sdk: gradio
sdk_version: 6.16.0
app_file: app.py
pinned: false
license: mit
short_description: Five small models run a market drama; you are the financier
tags:
- track:wood
- sponsor:openbmb
- sponsor:nvidia
- sponsor:modal
- achievement:welltuned
- achievement:offbrand
- achievement:sharing
- achievement:fieldnotes
- thousand token wood
- off-brand
- well-tuned
- best demo
- best agent
- field notes
- sharing is caring
- minicpm
- nemotron
- modal
---

# 🍄 Thousand Token Wood

A small-model **finance drama** for the [Build Small Hackathon](https://huggingface.co/build-small-hackathon).
Five woodland creatures, each driven by a *different lab's* small model, trade goods
for pebbles, gossip, scheme, and panic. You are the **Patron of the Wood**, a shadow
financier: you lend, whisper tips, short the market, bribe, and broker alliances, while
a magistrate watches for cheats. The chart is the scoreboard; the story is the people.

## Watch and read
- 🎬 **Demo video**: [YouTube (100s)](https://youtu.be/ugFyFumUCgs)
- 📣 **Social post**: [on X](https://x.com/RealLesterLeong/status/2064904967391961483)
- 📓 **Field notes** (the build, in four parts):
  [1 — emergent market drama](https://huggingface.co/blog/build-small-hackathon/thousand-token-wood-sim) ·
  [2 — the five-lab council](https://huggingface.co/blog/build-small-hackathon/thousand-token-wood-sim-v2) ·
  [3 — the crash that vanished](https://huggingface.co/blog/build-small-hackathon/thousand-token-wood-sim-v3) ·
  [4 — the wood fights back](https://huggingface.co/blog/build-small-hackathon/thousand-token-wood-sim-v4) ·
  [Medium mirror](https://medium.com/@LesterLeong/thousand-token-wood-emergent-market-drama-from-3-billion-parameter-agents-22545d5982bf)
- 🧠 **Fine-tuned model** (one of the five minds): [AdmiralTaco/ttw-trader-0.5b](https://huggingface.co/AdmiralTaco/ttw-trader-0.5b)
- 📡 **Open agent traces** on the Hub: [dataset](https://huggingface.co/datasets/build-small-hackathon/thousand-token-wood-traces)

## Five labs, five minds
Every creature thinks with a different small model, so the council is a live argument
between four labs, all under the 32B cap and served with vLLM on **Modal**:

| Creature | Model | Lab |
|---|---|---|
| Oona (owl) | gpt-oss-20b | OpenAI |
| Bramble (squirrel) | MiniCPM3-4B | OpenBMB |
| Fenn (fox) | Nemotron-Mini-4B | NVIDIA |
| Mossback (tortoise) + Pip (mouse) | ttw-trader-0.5b (fine-tuned) | ours |

The 0.5B was distilled from cleaned traces of a 3B teacher (self-buy mistakes stripped),
so a model one-sixth the size trades *more* cleanly than its teacher: zero self-buys,
100% valid offers. gpt-oss-20b doubles as the Narrator that writes the wood's headlines.

## Why "small" is load-bearing
A living economy needs *many* agents thinking *many* times per turn. Frontier models are
too slow and costly for that. Small models are what make a real-time council of traders
feasible, and the heterogeneity (four labs) is what makes the drama: each model has its
own temperament, so the market is an emergent negotiation, not a script.

## How Modal powers the wood
The entire model layer runs on **Modal**, end to end:
- **Four vLLM engines, four Modal apps** — one per lab's model, each on its own GPU,
  invoked per turn with **per-engine batching** (every creature on the same engine shares
  one batched call). All apps **scale to zero** when idle, so the wood costs ~nothing
  between sessions and the whole hackathon ran on a fraction of the starter credits.
- **Fine-tuning on Modal too**: `AdmiralTaco/ttw-trader-0.5b` was LoRA-trained on a Modal
  A10G from cleaned traces of its 3B teacher, then merged and pushed to the Hub — build,
  train, eval, and serve, all on one platform.
- **Cold-start craft**: a warmup ping precedes recorded runs, the UI streams a "waking the
  wood" frame during the first cold call, and the attract reel plays instantly with zero
  GPU so the Space is never blank while engines wake.

## Sponsor models in the council
- **OpenBMB MiniCPM3-4B** is Bramble the squirrel — the council's scrappy haggler.
- **NVIDIA Nemotron-Mini-4B** is Fenn the fox — the opportunist who buys the dip.
- **OpenAI gpt-oss-20b** is Oona the owl *and* the Wood Street Journal narrator.
Each model's temperament is visible in its creature's 💭 thoughts and raw JSON (both
exposed in the UI), so you can watch four labs' small models argue a market into being.

## What you can do (you are the Patron)
- **Power moves**: lend pebbles at interest, whisper a tip, short a good, bribe a creature,
  broker an alliance, or fund a corner. Every move costs you purse, reputation, or heat,
  and the creatures react.
- **The information war**: a tip can be *true* (real foresight of the next Wood Legend, your
  insider edge) or *false* (bait). Profit from an insider tip and your **heat** rises;
  cross the line and **Magistrate Heron** opens an investigation that ends in a fine, frozen
  pebbles, or exile.
- **Grudges and cartels**: creatures remember how you and others treated them. Short a
  creature's crop and it turns hostile and refuses your loans; broker an alliance and the
  partners stop undercutting each other.
- **The wood fights back**: burn the same creatures twice and they wise up. Wary holders
  stand against your next crash and blunt it, and they testify to the magistrate, so a
  gambit you keep repeating pays less every time and brings Heron faster. A live exposure
  meter shows your expected payoff collapsing as the wood sours on you.
- **Leverage and ruin**: loans come due, shorts get margin-called, and a creature that
  cannot pay is ruined and banished (it returns a chapter later).
- **Tempt Fate**: draw a *Wood Legend*, a famous market mania reskinned as woodland folklore
  (Tulip Mania, the South Sea Bubble, the 1929 bank runs, the 2020 toilet-paper scramble,
  the Hunt silver corner, the Dust Bowl), each carrying the real history under the fur.
- **Read their minds**: every creature's private reasoning is shown and shared as open traces.

## How it works
- `ttw/` is the engine: a deterministic double-auction market (`market.py`), the turn loop
  (`sim.py`), the per-creature multi-model policy (`agents.py`, `council.py`, `llm.py`), the
  Patron power moves (`moves.py`), relationships (`relationships.py`), the Wood Legends deck
  (`events.py`), the narrator and chapters (`narrator.py`, `chapters.py`), and a tolerant
  JSON parse-and-repair layer (`actions.py`) so no small model can break the simulation.
- `serve_council.py` serves each model as its own Modal vLLM engine.
- `app.py` is this Gradio app, the Wood Street Terminal console.

## Run it yourself
```bash
pip install -r requirements.txt
# No GPU, dummy agents (for trying the UI):
TTW_DUMMY=1 python app.py
# The multi-model council on Modal (deploy each engine, then enable the council):
TTW_APP_NAME=ttw-serve-gptoss   TTW_MODEL=openai/gpt-oss-20b           TTW_CUDA_DEVEL=1 TTW_REASONING_EFFORT=low python -m modal deploy serve_council.py
TTW_APP_NAME=ttw-serve-minicpm  TTW_MODEL=openbmb/MiniCPM3-4B          TTW_CUDA_DEVEL=1 TTW_TRUST_REMOTE=1 python -m modal deploy serve_council.py
TTW_APP_NAME=ttw-serve-nemotron TTW_MODEL=nvidia/Nemotron-Mini-4B-Instruct TTW_CUDA_DEVEL=1              python -m modal deploy serve_council.py
TTW_APP_NAME=ttw-serve-qwen     TTW_MODEL=AdmiralTaco/ttw-trader-0.5b  TTW_CUDA_DEVEL=1                  python -m modal deploy serve_council.py
TTW_COUNCIL=1 TTW_NARRATOR=1 python app.py
```

Built for the Build Small Hackathon, 2026. Small models, big adventures.