Homunculi chatbots
Small language models trained from scratch on our own machines in Berlin, ranked best first on an easy 40-question test (all questions and answers). Each chat opens in its own window. Small score differences are not a reliable overall ranking.
What the new v5 adds
V5 is an experimental maths/science specialist. It now routes supported English questions to its trained formats, shows the model's answer, and checks it against independent rules. Corrections are labelled separately.
English chat test: 16/40, fifth in this small test. The questions and grading rules were unchanged. All 40 answers used the legacy fact/link route; this test did not exercise v5's new calculation routes. See every answer and reviewed grade or the v5 run data.
| Sealed exact answers, 11 truth/prediction topics | 1,110 / 2,816 · 39.4% |
|---|---|
| Accepted by each topic's rules | 1,261 / 2,816 · 44.8% |
| Sealed maths, all forms | 191 / 256 · 74.6% |
| Worked arithmetic subset | 36 / 36 |
| Sampled trained-fact recall | 969 / 1,000 · 96.9% |
These research questions used dense prompt formats and a different dataset. They are not substitutes for the English chat score, and the other models did not take this sealed test. World handoff checks permit 5% numerical error; exact accuracy is shown separately. Several numerical physics predictions remain only 0–1.6% exact.
Choose by task: Classic for conversational text; Haishool v4b for established facts and graph connections; v5 for experimenting with checked maths and science. The links explorer is another interface to v4b's graph. None of these small experiments is consistently reliable.