MiniMax H3 runs with 4B/8B text encoder instead of 32B

A user replaced MiniMax H3's 32B text encoder (15.7 GB) with a Qwen3-VL 4B or 8B plus a learned map, keeping the same DiT, VAEs, and sampler. Video compares 32B vs 8B projected outputs.
1 source
Daily brief
Get tomorrow's AI brief in your inbox
More stories today
- Enterprise AI agents limited by messy documents
- Seinfeld AI video shows George in GTA 6 using Minimax H3
- Claude Code adds unrequested corrections to spec
- Ethan Mollick: AI impact research must address older-model limits
- Hobbyist trains 1.2B game music generator on single H100