SmolLM2

Single-GPU Small Model Beats Peers, but Results Still Need Replication

ConeML trained a 1.17B sparse MoE on one RTX 5090 and led same-size models in fact-chain tests, but its self-reported results lack validation.