MiniMax H3 reportedly uses Qwen3-VL-32B text encoder, split Transformer

Community analysis of a ComfyUI pull request suggests MiniMax H3 uses Qwen3-VL-32B (50 layers) as its text encoder and features a split Transformer architecture. Details emerged from Comfy-Org/ComfyUI PR #15210.
1 source
Daily brief
Get tomorrow's AI brief in your inbox
More stories today
- Mollick: ChatGPT Work, Claude Cowork should explain choices like a PM
- Deedy Das: AI-written prose can evade detection
- Podcast discusses the risks of AI-driven team velocity
- Alex Kantrowitz examines why Big Tech is falling behind in AI
- Stanford researchers change how AI agents access files