Minimax H3 text encoder appears online ahead of model release

Files for the unreleased Minimax H3's text encoder, Qwen3-VL-32B-Instruct-layer50_bf16, appeared online in 51.5 GB bf16 and 26.7 GB int8 forms. The poster says using smaller quantized text encoders can reduce prompt adherence, so larger versions may be better for accuracy.