Pruned BF16 MiniMax H3 Models Released, 40% Smaller
marres · reddit · 2026-08-05
Comfy-Org released pruned BF16 checkpoints for MiniMax H3, including FL2VA and REF2VA, each about 40.2GB, down from 66.3GB full BF16, removing 40% of parameters via precomputed AdaLN table without pruning main transformer blocks. A new quality-first option for BF16 users.
Related event: MiniMax H3 Pruned BF16 Model Released with 40% Size Reduction(2 posts)→
More from Models
- Qwen Devs Tease Upcoming 27B Model with 'New Level of Capability' — cedric_chee · 2026-08-05
- Maple-Preview: Open-Source Ternary-Weight LLM Hits 200+ tokens/s on Mac Mini — garrytan · 2026-08-05
- Reddit Discussion: Seeking Coding Finetunes Better Than Qwen 27B — Borkato · 2026-08-05
- DeepSeek V4 Flash Surfaces on OpenRouter: 284B Total Params, 1M Context — MikePFrank · 2026-08-05
- Extreme Quantization: DeepSeek-V4-Flash Crushed to 54GB GGUF — giveen · 2026-08-05
- Netizen Tests Minimax H3: Generates Flawless Code with a Single Prompt — CSProfKGD · 2026-08-05