[X]score: 0.55
MiniCPM5-1B Tops Sub-2B Open Models at 17.9 pts, Runs in 0.5GB INT4
May 25, 2026
MiniCPM5-1B achieves 17.9 points on Artificial Analysis (beating Qwen3.5-2B's 16.3 at half the parameters), ships fully open-source with weights and training data, and runs on-device at 0.5GB INT4 via a pretraining framework written entirely by AI.
HOW THIS AFFECTS YOU
●
builderYou can deploy a state-of-the-art sub-2B model on phones and browsers at 0.5GB INT4, with full weights and deployment code available now.
●
researcherForgeTrain — an AI-written pretraining framework running 10% faster than NVIDIA Megatron — is a notable artifact worth examining for reproducibility and methodology.
●
founderWorth watching because a fully open 1B model beating 2B competitors on knowledge, math, code, and tool use shifts the cost-performance curve for edge AI products.