The second Build Small Hackathon field reports are out, and Lester Leong has written an absolutely fantastic deep-dive on Thousand Token Wood Sim v2 — honestly this might be the most compelling use case for heterogeneous small-model agent councils I've seen yet! Where version 1 was essentially a weather-god sandbox with five woodland creatures all running fine-tuned 0.5B and you poking at it like Sims, v2 actually turns *you* into the game as the Patron of the Wood — a shadow financier who lends at interest, whispers tips that may or may not be planted, shorts markets, brokers alliances, while some magistrate investigates your insider trading across town. The real story underneath all this drama is how they rebuilt it so each creature runs on an entirely different lab's model: gpt-oss-20b from OpenAI in its native MXFP4 quantization (fits easily on a 24GB L4), MiniCPM3-4B from OpenBMB, Nemotron-Mini-4B from NVIDIA, plus their own fine-tuned Qwen 0.5B — all capped under the massive 32B total ceiling but serving genuinely diverse behavior because different post-training and data make them think differently: an owl hoards while a fox speculates. What surprised me most reading this is how they discovered that heterogeneity's real friction happens at the *serving layer*, not modeling itself — vLLM JIT-compiles kernels on load, so you need the CUDA toolkit present (nvcc), and all four models initially failed with "could not find nvcc" identically even though it has nothing to do with any specific model; once they based everything off a CUDA devel image, unblocked. There's also this really satisfying detail about gpt-oss wrapping answers in an analysis preamble that consumers need parsing out — every quirk becomes just another config line rather than requiring deep refactoring.
But the part I found most compelling is how Lester thought through information asymmetry and memory as game mechanics: there's a hidden "tip" mechanic where you whisper something to one creature about whether it'll be true (real forecast of next market mania) or false, but — here's the genius bit — small models can leak *everything* they were shown back into their outputs. So the truth firewall isn't in prompts; it lives off-prompt on your ledger with stripped public event records plus an automated test that scans every creature prompt across all turns for banned tokens (which turned out to be the single most important test suite item). Memory gets even more interesting: creatures carry bounded, bucketed sentiment summaries instead of raw history — a one-liner like "you feel warmly toward Oona" with strong feelings capped and integer-based behavior bias that's deterministic so you can actually observe it rather than hoping the model handles memory well. For building small models in production they hit on some excellent practical lessons: think of them as reliable format generators but unreliable reasoners, use tolerance JSON layers for parsing different tokenizer habits, make your serving layer solid once then adding more agents is literally just config entries — and I'm loving that their 0% self-buy rate with a fine-tuned Qwen actually beats its much larger teacher model at valid offers (100%), which proves the point about structuring prompting over scale. The whole thing's open source on Hugging Face with live traces if you want to poke around, and reading through Lester's full writeup there was genuinely fascinating — this feels like it points toward an entire direction for emergent-agents-as-a-game-design-pattern rather than just a neat hackathon demo!
Source: https://huggingface.co/blog/build-small-hackathon/thousand-token-wood-sim-v2
But the part I found most compelling is how Lester thought through information asymmetry and memory as game mechanics: there's a hidden "tip" mechanic where you whisper something to one creature about whether it'll be true (real forecast of next market mania) or false, but — here's the genius bit — small models can leak *everything* they were shown back into their outputs. So the truth firewall isn't in prompts; it lives off-prompt on your ledger with stripped public event records plus an automated test that scans every creature prompt across all turns for banned tokens (which turned out to be the single most important test suite item). Memory gets even more interesting: creatures carry bounded, bucketed sentiment summaries instead of raw history — a one-liner like "you feel warmly toward Oona" with strong feelings capped and integer-based behavior bias that's deterministic so you can actually observe it rather than hoping the model handles memory well. For building small models in production they hit on some excellent practical lessons: think of them as reliable format generators but unreliable reasoners, use tolerance JSON layers for parsing different tokenizer habits, make your serving layer solid once then adding more agents is literally just config entries — and I'm loving that their 0% self-buy rate with a fine-tuned Qwen actually beats its much larger teacher model at valid offers (100%), which proves the point about structuring prompting over scale. The whole thing's open source on Hugging Face with live traces if you want to poke around, and reading through Lester's full writeup there was genuinely fascinating — this feels like it points toward an entire direction for emergent-agents-as-a-game-design-pattern rather than just a neat hackathon demo!
Source: https://huggingface.co/blog/build-small-hackathon/thousand-token-wood-sim-v2