DmitryDB/MiniMax-H3-10Eros-Max-DT-sQKV

10Eros Max MiniMax-H3 Dynamic Time / Separate QKV

Patch-required FL2VA checkpoint for the 10Eros_Max fine-tune of MiniMaxAI/MiniMax-H3.

This is the dynamic-time, physically separate-Q/K/V counterpart of the stock-compatible 10Eros_Max INT8 checkpoint. It retains the original FP32 time MLP and dynamic rank-16 AdaLN path instead of using an adaln_t_table.

A ComfyUI core patch is required. Stock ComfyUI cannot load this file.

The compatible patch and its base-model documentation are maintained in MiniMax-H3-DynTime-sQKV. Download ComfyUI-MiniMax-H3-DT-sQKV.patch from that repository. The patch is deliberately not duplicated here.

This is a community conversion, not an official MiniMax or ComfyOrg release.

Download

Profile	Direct download	File size	Intended hardware
DT-sQKV INT8 ConvRot	FL2VA	20.999 GiB	24 GB · RTX 30/40. Quality-oriented experimental profile; workflow peak VRAM may still require offload.

Layout

Property	This checkpoint	Stock-compatible 10Eros_Max quants
Time conditioning	Original FP32 time_embedder MLP plus dynamic rank-16 AdaLN basis/mean	Rank-16 FP32, 4,097-point adaln_t_table
Attention input projections	Physical q_proj, k_proj, and v_proj tensors	Fused qkv_proj = cat(Q,K,V)
Main-matrix policy	170 INT8 ConvRot + 30 BF16 semantic matrices	Depends on selected stock profile
Physical quantized modules	270 INT8 modules because Q/K/V are separate	170 INT8 modules in the matching fused profile
Token refiner	Eight BF16 matrices	Eight BF16 matrices in the matching INT8 profile
ComfyUI compatibility	Core patch required	Stock ComfyUI

All 50 MiniMax-H3 transformer blocks remain present. This layout preserves the runtime time-conditioning calculation and avoids fusing Q, K, and V. It does not add model capacity; it changes how the original computation is represented and loaded.

Checkpoint size is not full-workflow peak VRAM. Resolution, frame count, attention backend, the Qwen3-VL text encoder, VAEs, and offload settings also consume memory.

Fine-tune conversion

The supplied full-width BF16 checkpoint was compared tensor-by-tensor with the local original MiniMax-H3 FL2VA source:

• all 535 expected tensors and all 50 blocks were present;
• shapes and dtypes matched the full-width architecture;
• no NaN or Inf values were found;
• 503 tensors were bit-identical to the original source;
• the intentional fine-tune consisted of 32 modified fused-QKV tensors in blocks 0–31.

For this checkpoint, each modified fused-QKV tensor was quantized with the same tensorwise INT8 ConvRot/Hadamard recipe used by the matching base profile, then split on exact Q/K/V row boundaries. Q, K, and V use independent per-row FP32 scales and quantization configuration tensors. A control comparison confirmed that the split payload is bit-identical to independently quantizing the three projections with the current quantizer.

Every tensor outside the permitted Q/K/V weight, scale, and configuration ranges remains byte-identical to the corresponding base DT-sQKV checkpoint.

Validation status

The file passed an independent byte-range audit:

1. file size, safetensors header, key inventory, shapes, and dtypes match the corresponding DT-sQKV reference checkpoint;
2. every byte outside Q/K/V ranges in blocks 0–31 remains identical to the reference;
3. all 288 replaced tensors match the separately generated 10Eros_Max payload;
4. modified weights differ from the base reference as expected.

Conversion and audit were CPU-only. No prompt-to-decoded-video perceptual A/B result is claimed yet. This validation proves container and conversion correctness, not subjective image or motion quality.

Installation

1. Back up or commit your ComfyUI core changes.
2. Download the linked core patch and verify it with git apply --check from the ComfyUI repository root before applying it.
3. Place the checkpoint in ComfyUI/models/diffusion_models/.
4. Restart ComfyUI and use the normal MiniMax-H3 FL2VA workflow.

The Qwen3-VL text encoder and the MiniMax-H3 video/audio VAEs are required but are not duplicated here.

License and attribution

Use is subject to the included MiniMax-H3 Community License Agreement. The base model is by MiniMax. The BF16 fine-tune was supplied under the name 10Eros_Max; this repository provides a quantized conversion and does not claim authorship of the fine-tune itself.