One GPU.
Video with sound.

Test an idea before making the final film.

A short scene can help explore an introduction, a food story or a camera move. Our samples cover a Singapore hawker setting, an English presenter, a Mandarin-speaking chef and a synthetic city flyover. The scenes, people and audio are AI-generated with MiniMax H3 through FastH3 V2.

Our takeaway: this is useful for trying a visual direction, then reviewing it. Keep precise lettering, logos and factual claims in an editing step you control. Speech, lip movement, hands and continuity still need close review before using an output in finished work.

This is recorded engineering evidence. Video generation is a workflow to scope and validate with us; these results do not establish that it is included in the standard workstation bundle.

Four samples. Their actual timings.

MiniMax H3 samples are unavailable in the US, EU, UK and South Korea, or where the viewing region cannot be established. The timings and test notes remain available. Regional availability uses IP Geolocation by DB-IP.

Use the homepage player, or open each MP4 below. These are complete selected outputs at original speed, re-encoded for web delivery with generated stereo audio. Generation time means prompt submission to the saved MP4, with models already warm. Each row is one observed output, not a median.

The city file is resized to 1792 × 1024 for web playback. The other three retain 1344 × 768. All four play at 24 frames per second. They show the generated result, not a real-time recording of the GPU generating it.

What to look and listen for.

• Hawker story: a hall, a cook at the wok and a shared meal. Inspect the transitions and signs; lettering in the generated stalls is unreliable.
• English presenter: a synthetic person in an office, prompted to welcome a new teammate. Listen to the speech and compare it with mouth movement.
• Mandarin chef: a synthetic cook presenting dumplings. Review the spoken words, hands and food detail together.
• City flyover: a synthetic waterfront and moving camera. Look at buildings, boats and fine detail. It is not footage of a real city.

The scene descriptions are viewing notes, not verbatim transcripts. We have not published a scored speech, lip-sync or human-preference evaluation.

How long does generation take?

The four prompts were a coffee product shot, an English presenter, a city flyover and a Mandarin chef, with fixed seeds. Times include text encoding, denoising, video/audio decoding and saving the MP4. They exclude downloading the models and the warm-up run. These small prompt sets do not give a statistical confidence interval or a guaranteed turnaround time.

Download the measured timings and sample counts (CSV). The medians were recalculated from the successful measured jobs; warm-ups and cached restyling jobs are excluded.

Cold start matters: after unloading the models, one 5.2-second, 1344 × 768 clip took 147.3 seconds. Do not expect the warm median for a first request.

Queued work is different from simultaneous work: eight 5.2-second clips queued back to back had a median of 56.2 seconds each. That was sequential generation, not eight simultaneous video jobs.

The configuration behind the numbers.

• One RTX 5090 cloud GPU, 32 GB VRAM, with a 125 GB host-memory cap.
• FastH3 V2 8-step, pruned INT8 ConvRot weights, using ComfyUI’s native workflow.
• Qwen3-VL-32B NVFP4-AWQ text encoder, FP16 video decoder and FP32 audio decoder.
• Eight sampling steps, CFG 1, video/audio shifts 10/3, and VSA sparse attention keeping 10% from 20% of the schedule.
• NVIDIA driver 580.126.09, CUDA 13.0 and PyTorch 2.14.0+cu130, as recorded on 18 September 2026.

The ComfyUI checkpoint listing identifies the model and encoder variants. The FastH3 V2 model card describes text-to-video-and-audio generation and its scope. Our timings are from our own single-GPU tests; they are not the project’s B200 measurements.

What changes the buying decision?

RAM as well as VRAM. The core tests used most of the card’s 32 GB VRAM. The test host had 125 GB RAM; an observed 4K upscaling workflow reached about 111 GB of process RAM. These runs do not qualify the TQ–32’s 64 GB configuration for every video workflow.

Native resolution and upscaling are different. A 3200 × 1792, 5.2-second render completed in 557.4 seconds; the next tested size, 3840 × 2176, ran out of GPU memory. That is an observed resolution ladder for one prompt and seed, not a universal maximum. A separate 3840 × 2160 output was upscaled from 1344 × 768 and must not be described as native 4K.

Quantisation is not automatically a speed-up. The tested community Q4_K_M workflow took a median of 155.9 seconds for a 5.2-second 1344 × 768 clip, versus 56.6 seconds for this INT8 workflow. Runtime kernels and model loading differed; this does not establish that all 4-bit models are slower.

A faster render is not a quality score. No matched human evaluation establishes equivalence to the base model. The samples are selected examples, and visual artifacts, speech mistakes or continuity errors can still appear.

One workload at a time in this test. We did not measure Qwen coding sessions running alongside video generation. Integration, heat, acoustics and sustained reliability of the assembled workstation remain separate qualification work.

Check the intended use.

FastH3 inherits the MiniMax H3 Community License (upstream copy). Its terms cover generated outputs as well as the model, including territory restrictions, attribution and disclosure of AI-generated content. Do not assume unrestricted use or distribution from the availability of downloadable weights.

For an agreed local video workflow in Singapore, discuss your intended use with our team. For the separate coding-agent measurements, read the Qwen / RTX 5090 test notes.