MiniMax H3 community experiments with smaller text encoders

A Reddit user replaced MiniMax H3's 32B text encoder with a 4B or 8B Qwen3-VL, using a learned projection to map hidden states, leaving the DiT untouched. Other posts compare Turbo LoRA steps and Sol-Attn for quality and speed.