These are real language models running entirely in your browser — tiny bees trained to speak buzz. They know almost nothing, and what they know is buzz. Pick a bee, say something, see what a few million parameters of bee sounds like. The right panel is a labeled simulation of swarm-verification — no real model there, just the idea drawn.
bees have nearly no memory — each reply sees only the last exchange or two. Mood bees answer with a [mood] tag; they hold grudges, notice kindness, and know flowers.
A concept sketch of swarm-verification: scouts inspect a claim and vote YES / NO / ABSTAIN. Contradictory scouts cancel each other (cross-inhibition); the hive commits only past quorum, and a split hive is an honest answer. When the real verifier eval lands, this panel replays real fixtures.
BeeLLM models by schizophyllume (Apache-2.0, huggingface.co/schizophyllume/BeeLLM); queen Mellifera by Sauers (huggingface.co/Sauers/BeeLLM-Mellifera). Inference: wllama — llama.cpp compiled to WASM. Everything runs on your machine; nothing you say to a bee leaves this page. The 12M queen loads from the Hugging Face CDN only when you summon her.