10Eros-Max-beta2-NVFP4

10Eros-Max beta2 — NVFP4 (ComfyUI native)

40.22GB BF16 → 12.53GB NVFP4. Runs on a single 16GB card.

NVFP4 quantization of TenStrip/10Eros-Max (beta2_pruned) — the MiniMax-H3 graft experiment carrying character from LTX-2.3 / Wan 2.2 / Krea 2. The author ships BF16 only (40GB); this is the same model at less than a third of the size.

Recipe (the part worth copying)

Instead of inventing a precision layout, copy the vessel: read which 200 layers the official Comfy-Org H3 NVFP4 quantizes (the ones carrying .comfy_quant), quantize exactly those, and pass the other 332 tensors through untouched. Key structure verified against the official file: 0 missing / 0 extra / 0 shape mismatches. Output lands at the same 12.53GB.

On the author's attn_k warning ("touching K silently degrades audio"): that concern is about weight surgery. For quantization, the official H3 NVFP4 already quantizes the fused qkv_proj in production — this repo follows that proven layout. Bake script included (bake_eros_nvfp4.py), ~24s on one RTX PRO 2000 Blackwell.

Usage (ComfyUI)

Drop into models/diffusion_models/, load with UNETLoader, and reuse your existing MiniMax-H3 workflow as-is: CLIPLoader type minimax, H3 video/audio VAE, MiniMaxH3ImageToVideo, MiniMaxH3SigmaShift (12.0 / 3.0), res_multistep.

Measured: 480×864, 145 frames, 20 steps, with audio — 16m39s on a single RTX PRO 2000 Blackwell (16GB), dynamic VRAM staging ~12GB.

Licenses

MiniMax-H3 community license applies. Per the upstream author, the community licenses of the character donors (LTX-2.3 / Wan 2.2 / Krea 2) also apply to the transferred character portions. Quantization changes none of this.