What FastH3 Omni adds
Hao AI Lab's October 20 release describes an eight-evaluation distilled MiniMax H3 model for reference-guided video and audio. Inputs can include a prompt, images, video clips and a voice reference. The team provides weights and serving recipes, followed by local-hardware guidance on October 21.
Separate server speed from desktop speed
The release reports 5.6–11.9 times lower end-to-end latency across 27 matched workloads on eight GB200 GPUs, compared with Base's lossless-policy serving recipe. That is a comparison of specified serving configurations. It is not a promise that every desktop can generate a video faster than its playback duration.
Read the local timing footnotes
For a one-photo, five-second 480p task, the follow-up reports about 62 seconds on an RTX 5090, 192 seconds on one DGX Spark, 16.9 minutes on an RTX 3090 and 19.5 minutes on an M5 Max. The Mac value is a single run, while the 3090 time excludes roughly seven minutes of model loading. Different numerical variants can produce different outputs.
A reference image changes the workload
Our suggested comparison begins with one original reference image and a five-second brief. Keep resolution, duration and requested action constant. Record cold loading, prompt preparation, generation and export separately. Then add a voice reference you have permission to use and repeat the run. A reference-conditioned request does more work than plain text-to-video, so do not compare their timings as if they were identical tasks.
Judge usable clips, not only finished files
For each output, inspect whether the subject retains its identity, whether the requested movement is present and whether speech matches the intended words. Listen without watching once to catch audio artifacts hidden by attractive imagery. Keep failed takes and record the reason for rejection. The useful production measure includes the time required to produce an accepted clip, including retries and editing. This is our proposed evaluation, not a completed FyreLinkz benchmark.
Choose the recipe for the actual machine
Use the project's device-specific recipe and verify the linked checkpoint's terms before downloading. Check available disk space, host memory and GPU memory independently. Keep first tests short and avoid assuming that a community result shares your settings. Our video cost comparison and calculator can help track spend per accepted minute when comparing a local workflow with hosted generation. A local run still has hardware and time costs even when there is no per-call API bill.
Frequently asked questions
Does FastH3 Omni generate audio too?
Yes. The project describes finished video with synchronized sound and supports image, video and voice references.
Does the 12× speed claim apply to an RTX laptop?
No. The stated comparison uses eight GB200 GPUs and specified serving recipes. Local results have separate settings and limitations.
Are the published desktop timings FyreLinkz measurements?
No. They are Hao AI Lab's reported results. FyreLinkz has not independently reproduced them.
Sources & further reading
Primary sources checked Oct 25, 2026. Vendor statements are attributed; editorial advice is our own.
- 1
- 2
Help us keep this useful. Report a factual correction →
