跪拜 Guibai
← All articles
Frontend · Backend · Artificial Intelligence

MiniMax H3 Lands on a 5060 Ti: Native Audio-Video Generation Squeezed into 12.5 GB

By 雪隐_上班了 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

A 33B video model with native audio running on a single consumer GPU removes the multi-model pipeline that made AI filmmaking brittle — no more desynchronized lip-sync, mismatched soundscapes, or Frankenstein editing across four different tools. The NVFP4 quantization technique, which surgically preserves full precision for error-sensitive AdaLN layers while compressing attention and MLP blocks, sets a pattern for squeezing large diffusion models onto desktop hardware.

Summary

MiniMax open-sourced H3, a 33B full-modality model that generates video with synchronized dialogue, sound effects, and music — all produced natively during the diffusion process rather than stitched together from separate models. The official weights demand over 120 GB of VRAM, but a community NVFP4 quantization, built on Comfy-Org's pruned version that collapsed 13B AdaLN parameters into a 40-million-entry lookup table, shrinks the model to 12.5 GB. It runs on an RTX 5060 Ti 16 GB with dynamic offloading, albeit slowly.

The model comes in two variants: fl2va for text-to-video, image-to-video, and frame interpolation with infinite sequels, and ref2va for reference-based generation that locks character identity, style, and voice across up to nine reference images. Prompting requires structured storyboard scripts with timestamped shots, explicit dialogue tags, and dedicated sections for ambient sound and music — unstructured prose produces incoherent output.

Deployment involves upgrading ComfyUI to 0.30.0, placing roughly 34 GB of model files across diffusion_models, text_encoders, and vae directories, and loading official workflow templates with a single node swap to point at the NVFP4 weights. Frame counts must follow the formula 17n+5.

Takeaways
MiniMax H3 generates video and native stereo audio together in one diffusion process, not as separate post-production steps.
The model has two variants: fl2va handles text-to-video, image-to-video, and frame interpolation; ref2va locks character identity and style using up to nine reference images, three reference videos, and three reference audio clips.
Official full-precision weights require over 120 GB VRAM; Comfy-Org's int8 version drops that to 20 GB, and a community NVFP4 quant further reduces it to 12.5 GB.
The NVFP4 quant works only on Blackwell architecture GPUs (RTX 50 series); older cards must use the int8 version.
Comfy-Org's pruned version identified that 13B AdaLN parameters depend only on the timestep and collapsed them into an 8-dimensional lookup table, a 326x compression that keeps the error-sensitive modulation layer at full precision.
Frame counts must follow the formula 17n+5; for example, a 5-second clip at 24fps requires exactly 125 frames.
Prompts must use structured storyboard syntax with timestamped shots, explicit dialogue tags like `<d>[English] text</d>`, and dedicated sections for overall_soundscape and non_diegetic_music.
The full model file set occupies about 34 GB on disk, split across diffusion_models, text_encoders, and vae directories.
VAE files are deliberately kept unquantized because they run only once per generation and quantization errors would be directly visible in output.
Deployment requires ComfyUI 0.30.0 or higher, with the sampler set to res_multistep and CLIPLoader type set to minimax.
Conclusions

Collapsing AdaLN into a lookup table before quantizing the rest of the model is a transferable strategy: identify the small, error-sensitive subsystem, keep it at full precision, and compress everything else aggressively.

The 17n+5 frame constraint suggests the model's temporal layers operate on a fixed stride that doesn't divide evenly into common frame counts, a detail that will trip up anyone who skips the documentation.

Requiring structured storyboard scripts rather than natural-language prompts means H3 is effectively a rendering engine for a formal intermediate representation, not a conversational creative tool — the user supplies direction, not description.

Native audio generation inside the diffusion process eliminates the multi-model pipeline that has been the standard approach since the first open-source video models appeared, collapsing four separate inference steps into one.

Concepts & terms
AdaLN (Adaptive Layer Normalization)
A modulation mechanism in diffusion models that applies per-layer scaling coefficients to residual streams based on the current timestep. Errors in AdaLN parameters multiply across all layers, making it the riskiest part of a model to quantize.
NVFP4
A 4-bit floating-point quantization format supported by NVIDIA's Blackwell architecture (RTX 50 series). It offers higher throughput and lower memory usage than int8 while preserving more dynamic range than integer quantization at the same bit width.
H3-Context-IR
MiniMax's structured intermediate representation for video prompts, consisting of timestamped shots, explicit dialogue tags, camera movement specifications, and dedicated soundscape sections. The model was trained on this format and performs poorly with unstructured natural-language prompts.
Dynamic Offloading
A memory management technique where model layers are moved between GPU VRAM and system RAM during inference, allowing models larger than available VRAM to run at reduced speed.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗