The creator of Redis just brought one of the world's most powerful AI models to Mac, no NVIDIA GPU required

One developer writing C and Metal shaders generated a 22-frame video from a 33-billion-parameter AI model in 3.5 seconds. No NVIDIA GPU, no Python framework, no CUDA runtime. Salvatore Sanfilippo, best known by his handle antirez for creating Redis, released h3.c, a from-scratch inference engine that runs MiniMax H3 natively on Apple Silicon 1. The project reached the front page of Hacker News and drew hundreds of GitHub stars within hours 2.

MiniMax H3 is a 33-billion-parameter diffusion Transformer with 50 layers that takes text, images, video, and audio as input and generates video with native stereo audio at resolutions up to 2K 3. An independent MLX port of the model is blunt about what it is and is not: H3 is a diffusion transformer that denoises video and audio latents jointly, conditioned by a frozen Qwen3-VL-32B text encoder, not a language model 4. Released as open weights, it shipped with inference tooling built for NVIDIA CUDA. AMD added Day 0 ROCm support for H3 on its Instinct GPU line 5. On Apple Silicon, the official ComfyUI Text-to-Video workflow fails outright: the video output is completely black, and audio encoding throws NaN errors. The Image-to-Video and Reference-to-Video workflows run correctly on the same machine 6.

What h3.c runs, and how fast

The engine handles prompt-to-video, prompt-to-audio, first and last frame conditioning, and reference-based generation end to end, all implemented in C with Metal compute shaders 1. On an M5 Max, the fastest validated configuration generates 22 frames at 512x512 resolution in approximately 3.5 seconds using four denoising passes. A 29-pass reference run on the same hardware takes 26.4 seconds 1. For comparison, the broken ComfyUI Text-to-Video run that produced a black screen took 359 seconds on an M4 Max 6. The hardware and pass counts differ, but the gap between six minutes to failure and 3.5 seconds to a working result is the software stack made visible in wall-clock time.

Those speed numbers are antirez's own measurements with no independent verification. The four-pass output achieves 0.556 structural similarity (SSIM) against the 29-pass reference on his fox test sequence, and 0.547 on a separate surfer test 1. SSIM scores how closely a reduced-quality output matches a reference frame by frame. The trade-off is a 7.5x speedup (3.5 seconds versus 26.4) for content that clears a minimum structural bar.

The memory math explains why a 33-billion-parameter model fits on a laptop at all. Approximately 13 billion of the model's parameters can be precomputed once and cached, dropping them from memory during generation 3. That leaves 20 billion active parameters. At BF16 precision, two bytes per parameter, 20 billion parameters occupy roughly 40 gigabytes. The independent MLX port confirms this with exact measurements: the DiT drops from 33.12 billion parameters (66.3 GB) to 20.11 billion (40.3 GB), a 25.3-gigabyte saving computed in 0.7 seconds 4.

Forty gigabytes for the core model fits inside the unified memory of Apple Silicon Macs with 64 gigabytes or more. The optimization that makes this model runnable on consumer hardware is a parameter-counting exercise that reduces the active footprint by 39%, not a new capability in Apple's Metal framework.

Why the official path broke

The ComfyUI failure is a software dependency problem, not a hardware limitation. The official Text-to-Video workflow runs H3's Qwen3-VL-32B text encoder through PyTorch's Apple MPS backend. The Image-to-Video and Reference-to-Video workflows work correctly on the same hardware 6. The issue reporter states the root cause has not been identified.

h3.c sidesteps the entire problem. It loads BF16 weights directly from safetensors files, implements every transformer block, text encoder pass, and VAE decoder in C, and computes through Metal shaders with no intermediate framework performing data-type conversions 1.

The CUDA dependency that shapes the AI hardware market sits in the inference software layer, not in the model architecture. Each platform that now runs H3 required its own implementation, but none required the model to change:

• NVIDIA CUDA: the reference path shipped with the model
• AMD ROCm: Day 0 support on Instinct GPUs
• Apple MLX: community-developed ports, validated against the diffusers reference
• Apple Metal, native: h3.c, written from scratch in C by one developer

MiniMax acknowledged the project on social media, noting that open weights mean anyone can bring the model to new hardware 7. That acknowledgment confirms the project runs but is not an endorsement of output quality. For developers choosing hardware and investors pricing NVIDIA's compute advantage, the pattern across CUDA, ROCm, MLX, and now native Metal is consistent: the model ships once, and each platform builds its own path to run it.