A 1B Apache 2.0 model scoring 17.9 and using few tokens is relevant for on-device and low-cost inferenceRunning a trained model to get answers — the phase where AI is actually used, as opposed to trained.Full definition →.
Key takeaways · AI-distilled
MiniCPM5-1B leads every open-weights model at 1B parameters or below by 7.4 points, and beats Qwen3.5 2B (reasoning, 16.3) by 1.6 points with less than half the parameters. Unlike MiniCPM-V 4.6, it is text-only.
It used 12.6M output tokenThe chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.Full definition → to run the Intelligence Index, about 31x fewer than Qwen3.5 2B in reasoning mode (389M), though about 2.3x more than its predecessor MiniCPM-V 4.6 1.3B.
Its AA-Omniscience score of -1 is the best in its size class because it declines most questions instead of guessing; sub-2B peers that attempt answers land between -70 and -89 on the hallucinationWhen a model states something false with full confidence — inventing facts, citations, or APIs that don't exist.Full definition →-penalized index.
Terms in this piece · Glossary
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.
hallucination — When a model states something false with full confidence — inventing facts, citations, or APIs that don't exist.
