Tiny model frontier mapped: GLiNER 2.5 leads <350M, Kev family tops two brackets

multimodalart · x · 2026-09-22

multimodalart maps out the "tiny frontier" of small models: GLiNER 2.5 leads the <350M bracket, the Kev family tops both 350M-700M and 3B-10B, Decider 2B sits in the sub-3B sweet spot, and Jevfire's Qwen3.8 27B single-pass implementation is the best open-weights jev-like model.

Follow-up notes that jev-style models do particularly well on graduate-level knowledge questions, Python function choosing, tool calling and model routing. The latency-performance Pareto frontier also includes openvons' Qwen3-4B as a super-fast option and the now-classic diffusiongemma jev-like implementation.

Original post →

More from Fun

Fun channel →