Updated 2026-10-02. Last verified 2026-10-02 (separate from the article date).
Selected report: one aarch64 DGX Spark GB10, one GPU, 128 GB unified memory on a single pool, vLLM 0.26.0, measured vLLM-Omni code e1aa6eae75c460cd1893bc320546e81e66973831. FL2VA partition, T2VA, 960×576, 8 seconds requested, 24 fps, 50 steps. Client end-to-end 2169.4 s (36:09). Allocator high-water 97.7 GiB. This site did not reproduce the run.
Four components
Resident byte sizes are unknown. The diagram and the list name the same four parts.
H3 text encoderBF16
H3 omni transformer / DiTFP8 at measured code pin e1aa6eae75c460cd1893bc320546e81e66973831
Visual VAEBF16
Audio VAEBF16
H3 text encoder. Uses Qwen3-VL-32B. Vision understanding is part of this encoder, not a separate memory allocation. Precision: BF16. Placement: Resident in the unified pool, BF16, not offloaded. Resident bytes: unknown.
H3 omni transformer / DiT. 33B dense is an architecture label only. No 33B-plus-32B fit is calculated here. Precision: FP8 at measured code pin e1aa6eae75c460cd1893bc320546e81e66973831. Placement: FP8 at the measured code pin, in the same unified pool, not offloaded. Resident bytes: unknown.
Visual VAE. Separate visual latent autoencoder. Resident byte size is not measured on this page. Precision: BF16. Placement: BF16, tiled decode, in the same unified pool. Resident bytes: unknown.
Audio VAE. Separate audio latent autoencoder for stereo audio. Resident byte size is not measured on this page. Precision: BF16. Placement: BF16, in the same unified pool, byte size unmeasured. Resident bytes: unknown.
Selected workload and measurements
Hardware
One aarch64 DGX Spark (GB10), one unified-memory GPU
No CPU offload, no GPU offload, and no distributed layerwise offload. CPU and GPU are the same unified pool.
Tiling
Tiled VAE decode
Peak
97.7 GiB. Allocator high-water (max reserved). Excludes the CUDA context and non-PyTorch allocations. Not total system use and not free headroom.
Concurrency
One request at a time was reported. That is not a throughput guarantee.
Network
No multi-host network is required for this single host. Parallel degrees in the source config are 1.
Storage
135 GiB checkpoint storage per task partition is a disk requirement. One partition is resident. Disk size is not runtime offload evidence.
Client end-to-end. 2169.4 s (36:09)
Denoise. 2088 s
Per step. 42.61 s
MP4 mux. 4.84 s
Text encode. 0.25 s, transferred from a separate 10-step measurement at the same 960×576 / 8 s geometry. Do not add this into a new total.
VAE decode. 70 s, transferred from that same 10-step measurement. Do not add this into a new total.
Sampling. seed 1101, flow_shift 12, audio_flow_shift 3, aspect_ratio 16:9. The source requires an explicit aspect ratio even when width and height are set.
Topology. Single host. Tensor parallel, text-encoder parallel, USP, ring, and VAE patch parallel degrees are 1.
Sync timeout. 7200 seconds in the source configuration.
Output. Muxed MP4 with H.264 video and 32 kHz stereo AAC.
Platform memory label. 128 GB unified LPDDR5X, about 121 GiB usable, as stated by the source. This page does not subtract the allocator figure from that label.
FP8 scope. At the measured code pin, online FP8 quantizes the DiT only. The source describes that DiT change as 62 GiB to about 31 GiB. The encoder and both VAEs stay BF16. Later revisions change what the FP8 flag covers. The 62 GiB and 31 GiB figures are source weight notes, not a fit total.
Document versus code. Qualified document pin d81777f297811ceb68bd991b6defd9f46fc9162e states the measured code pin e1aa6eae75c460cd1893bc320546e81e66973831 and vLLM 0.26.0.
Unknowns
Frame count was not reported. It is not calculated from 8 seconds and 24 fps.
Measured checkpoint revision was not stated by the runtime report.
The allocator high-water is not total pool use and is not free headroom.
Stage times do not sum to end-to-end. Text encode and VAE decode on the 50-step row were transferred from a separate 10-step run.
Component resident sizes other than the source's DiT weight note are unknown.
A two-host driver version in the same document is not claimed as this one-host measurement.
License
MiniMax H3 Community License, dated 2026-08-02, source revision 42ed227ee7df40d41602854ae760620d6eb651fe. Applicable territory exclusions: EU, UK, Republic of Korea, USA. Commercial products or services above US$20M yearly revenue need separate prior written authorization. This page does not infer where you are and does not say your deployment is cleared.Full H3 license. Runtime Apache-2.0 does not replace that license. The Qwen3-VL-32B encoder remains under its own Apache-2.0 terms.
vLLM-Omni Apache-2.0revision e1aa6eae75c460cd1893bc320546e81e66973831. Runtime software license. Not the H3 model license. The Qwen3-VL-32B encoder is Apache-2.0 under its own terms.
Limits
Released local H3-Base generates at 768p. These two reports are not that default 768p description.
Hosted H3-Context-IR plus H3-Regenerate-2K is a separate service flow. This site does not call that API.
Ref2VA is a separate checkpoint partition. It is not the selected workload on either page.
The 33B dense Omni Transformer size is architecture context only. It is not added to the 32B encoder, and the experimental AdaLN cache is not subtracted.
NVMe or other disk capacity is not evidence of runtime offload.
A two-host note in the upstream document uses an experimental external environment launcher and a benchmark-only adaptation. Cross-host vllm serve is not a supported server command there. Four-host timing is unknown. This page has no two-host or four-host timings.
Rollback
Stop the attempted workload.
Restore the prior validated environment.
Recheck output and memory before another attempt.
No destructive deletion.
One aarch64 DGX Spark GB10, with one unified-memory GPU, finished an upstream vLLM-Omni T2VA request. The qualified document pin is d81777f297811ceb68bd991b6defd9f46fc9162e. That document says the measured code pin is e1aa6eae75c460cd1893bc320546e81e66973831 and vLLM is 0.26.0. Client end-to-end time is 2169.4 seconds (36:09). The allocator high-water is 97.7 GiB. This site did not reproduce the run. The allocator number is not total system use and it is not free headroom.
CPU and GPU share one pool. The source labels the machine as 128 GB unified memory, about 121 GiB usable. This page does not subtract 97.7 GiB from that label. There is no CPU offload, no GPU offload, and no distributed layerwise offload, because those paths would use the same pool.
What was selected
The checkpoint partition is FL2VA, task T2VA. Geometry is 960×576. The request asks for 8 seconds at 24 fps and 50 inference steps. Frame count was not reported, and this page does not calculate one. At the measured code pin, online FP8 covers the DiT only. The source describes that DiT weight change as 62 GiB to about 31 GiB. The Qwen3-VL encoder and both VAEs stay BF16. Later revisions change what the FP8 flag includes, so this scope stays tied to code pin e1aa6eae. Attention is cuDNN attention in eager mode. VAE decode is tiled.
Sampling in the source is seed 1101, flow shift 12, audio flow shift 3, and an explicit 16:9 aspect ratio even though width and height are set. Parallel degrees are 1. The source video sync timeout is 7200 seconds. The output is a muxed MP4: H.264 video plus 32 kHz stereo AAC. One request at a time was reported. That is not a throughput ceiling. Checkpoint storage is 135 GiB for the task partition, and one partition is resident. That disk figure is not runtime offload, and no SSD offload was measured.
Denoise on the selected row is 2088 seconds, about 42.61 seconds per step, and MP4 mux is 4.84 seconds. Text encode 0.25 seconds and VAE decode 70 seconds come from a separate 10-step measurement at the same 960×576 / 8 second geometry. They transfer onto the 50-step description because the source says those stages do not depend on step count. They are not added here into a new total. Stage times do not equal end-to-end time. The measured checkpoint revision was not stated. A driver version mentioned for a two-host qualification is not claimed for this one-host run.
Boundaries
The H3 text encoder uses Qwen3-VL-32B, including its vision understanding. The 33B dense transformer size is architecture context only. It is not added to the encoder, and AdaLN branches are not subtracted. Visual and audio VAE resident sizes, aside from the source DiT weight note, stay unknown.
The model license is the MiniMax H3 Community License dated 2026-08-02, revision 42ed227ee7df40d41602854ae760620d6eb651fe. Territory exclusions are the EU, the UK, the Republic of Korea, and the USA. Yearly commercial revenue above US$20M needs separate prior written authorization. vLLM-Omni Apache-2.0 is the runtime license, not the model license. The encoder’s Apache-2.0 note is separate. This page does not guess a reader’s location.
Local H3-Base at 768p and hosted Context-IR plus Regenerate-2K are different flows. Ref2VA is another partition and is outside this selected report. A two-host experiment uses an external environment launcher and a benchmark-only adaptation. Cross-host vllm serve is not a supported server command for that experiment. Four-host timing is unknown. Those cases have no timings on this page. The plain-text brief above is the planning copy and contains no launch command.