Vflash MiniMax H3 Ref2VA INT8-SGL

This repository publishes the full MiniMax H3 Ref2VA DiT converted to per-output-channel symmetric INT8 for LightX2V's int8-sgl runtime. It is a component checkpoint, not a standalone H3 model: use the official MiniMax H3 text encoder, video/audio VAEs, schedulers, processor, and safety controls.

Release identity

Item	Value
Base model	MiniMaxAI/MiniMax-H3
Base revision	42ed227ee7df40d41602854ae760620d6eb651fe
Source partition	transformer_ref
Converter	LightX2V a34806c0b35841e12982f33d27c7fbae135fad4e
Quantization	symmetric INT8 weights per output channel; dynamic INT8 activations
Runtime kernel	SGL INT8 GEMM
File size	34,023,112,968 bytes (31.69 GiB)
SHA-256	d10f8b6e6bf93bf1a76a7160901568be6394d6f7e4e26ed4a1a20bba03c6f246

The safetensors header contains 988 tensors: 350 INT8, 626 BF16 (including the corresponding scales), and 12 FP32 tensors. No layer was pruned.

Conversion

The checkpoint was produced with LightX2V's official converter contract:

python tools/convert/converter.py \
  --source /models/minimax-h3/transformer_ref \
  --output /models/quantized/h3-ref-int8-sgl \
  --output_name minimax_h3_ref_int8 \
  --model_type h3 \
  --quantized \
  --linear_type int8 \
  --device cpu \
  --single_file

Download the checkpoint separately from the official base model:

hf download rockstarengine/vflash \
  h3-ref-int8-sgl/minimax_h3_ref_int8.safetensors \
  --local-dir models/quantized

Set these LightX2V configuration fields:

{
  "dit_quantized": true,
  "dit_quant_scheme": "int8-sgl",
  "dit_quantized_ckpt": "/models/quantized/minimax_h3_ref_int8.safetensors"
}

The matching Vflash runtime profile and patches are in Hansimov/vflash.

Measured scope

All values below are local measurements, not upstream claims. Canvas is 928x512 and the output includes native stereo audio.

Hardware and contract	BF16 baseline	INT8 path	Speedup
RTX 4090 48 GB, Base50, 124 frames, hot E2E	600.833 s	419.795 s resident	1.431x
RTX 4090 48 GB, 4 updates, 362 frames, BF16+FA3 vs INT8+Sage2	252.424 s	178.833 s	1.412x
RTX 3080 20 GB, 4 updates, 362 frames, BF16+Sage2 vs INT8+Sage2	461.508 s	378.991 s	1.218x
RTX 4090 48 GB, Base40, 362 frames, BF16+Sage2 vs INT8+Sage2	1443.667 s	1060.874 s	1.361x

The RTX 3080 20 GB also completed a 40-update, 124-frame Ref2VA request with block offload in 787.364 seconds and 11,740 MiB observed physical VRAM.

Quality boundary

This is a fast-preview and constrained-memory checkpoint, not a lossless replacement for BF16. On the long Base40 comparison, sampled semantic blind reviews tied, but average frame-to-frame motion was 21.99% lower and audio RMS was 4.244 dB lower than BF16+Sage2. On the same-card Base50 comparison, SSIM was 0.741911 and the generated diffusion trajectory visibly diverged. Use BF16 for quality masters; qualify this checkpoint on your own motion, reference-identity, dialogue, and audio workloads.

License and access

MiniMax H3 and this model derivative are governed by the MiniMax H3 Community License Agreement, not by Vflash's Apache-2.0 source-code license. The license currently defines the Applicable Territory as worldwide except the European Union, United Kingdom, Republic of Korea, and United States, and imposes distribution, hosted-service, acceptable-use, and commercial conditions. Using or downloading this checkpoint means you are responsible for satisfying those terms.

Modified files carry this model card and NOTICE. The repository uses the same public, non-gated Hub access mode as the official MiniMaxAI/MiniMax-H3 repository; that access mode does not waive or replace the model license.