Phi-4 family (Microsoft)
14B main + Phi-4-mini (3.8B). Textbook-quality training data. Punches above weight. Reasoning strong for size.
Advertisement
Gemma 3 (Google)
1B, 4B, 12B, 27B. Multimodal + 128K context. Vision variants. Efficient on Google TPU + widely on consumer GPUs.
Advertisement
Qwen 3 (Alibaba)
0.5B to 32B. Strong multilingual + code. Both dense + MoE variants. Apache-2.0 friendly.