Updated 2026-10-02. Last verified 2026-10-02 (separate from the article date).
Selected report: one 24GB RTX 4090 D, SGLang, FL2VA partition, T2VA, kitchen_int8 DiT linears plus FlashAttention, layerwise DiT and text encoder, VAE left on device. Timed end-to-end 303.3 s and denoise 273.7 s. Peak GPU about 18 GB. 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 encoderUnspecified beyond the source statement that INT8 applies to DiT linears
H3 omni transformer / DiTkitchen_int8 linear layers
Visual VAENot described as INT8
Audio VAENot described as INT8
H3 text encoder. Uses Qwen3-VL-32B. Vision understanding is part of this encoder, not a separate memory allocation. Precision: Unspecified beyond the source statement that INT8 applies to DiT linears. Placement: Layerwise streaming with the DiT. Resident bytes: unknown.
H3 omni transformer / DiT. 33B dense is an architecture label only. No 33B-plus-32B fit is calculated here. Precision: kitchen_int8 linear layers. Placement: Layerwise streaming, 0 resident layers, prefetch 1. Resident bytes: unknown.
Visual VAE. Separate visual latent autoencoder. Resident byte size is not measured on this page. Precision: Not described as INT8. Placement: On device, tiled decode, excluded from layerwise offload. Resident bytes: unknown.
Audio VAE. Separate audio latent autoencoder for stereo audio. Resident byte size is not measured on this page. Precision: Not described as INT8. Placement: Not given a separate offload target in the report. Resident bytes: unknown.
Selected workload and measurements
Hardware
One 24GB RTX 4090 D
Runtime
SGLang. Installed SGLang version and commit were not reported.
Partition and task
FL2VA, t2va
Geometry
1344×768
Frames
107
Duration
not reported
Frames per second
not reported
NFE
20
Inference steps
not reported
Attention
FlashAttention (fa). Exact attention, selected.
Offload
Layerwise streaming of the DiT and text encoder. Zero resident DiT layers and prefetch 1, from the source command description. No measured SSD path.
Tiling
VAE stays on device with tiled decode. VAE is not a layerwise-offload component in the selected report.
Peak
about 18 GB. Upstream peak GPU for this report. The unit stays GB. It is not a GiB conversion and not a fit guarantee. Host RAM was not reported.
Concurrency
Not independently reported
Network
Single GPU. No multi-host network is part of this report.
Storage
No measured SSD or NVMe runtime path.
Selected timing: kitchen_int8 + FlashAttention, 303.3 s end-to-end, 273.7 s denoise, 1.34× versus BF16 + FlashAttention, PSNR 24.81 dB. Selected default. The source says this changes Linear numerics only. PSNR is versus BF16 + FlashAttention.
Comparison timings
These rows are not the selected config.
BF16 + FlashAttention: 405.6 s end-to-end, 370.2 s denoise, 1.00×, PSNR n/a dB. Baseline comparison. Not the selected config.
kitchen_int8 + sol_attn: 223.9 s end-to-end, 203.9 s denoise, 1.81×, PSNR 24.44 dB. Comparison only. Changes the attention algorithm.
kitchen_int8 + sage_attn: 174.9 s end-to-end, 154.2 s denoise, 2.32×, PSNR 23.51 dB. Comparison only. Changes the attention algorithm and the reported quality.
kitchen_int8 + Sage to Sol hybrid: 163.8 s end-to-end, 143.1 s denoise, 2.48×, PSNR 23.04 dB. Comparison only. Changes the attention algorithm.
Sampler. euler
torch.compile. disabled
Step caching. disabled
Method. Load, then warmup seed 0, then timed seed 42. Only the timed pass is reported, so setup is excluded.
Document pin. SGLang docs 2cd92d10b98a73de3952eafca915ae8954fcd375, section #rtx-4090-24-gb-single-gpu-run. The benchmark table is around source line 2105. That line is not an installed commit.
VAE offload warning. The source says putting the VAE decoder in layerwise offload re-streams about 9 GiB on each of 167 decode tiles. That placement is not the selected report. The selected peak remains about 18 GB.
Unknowns
Installed SGLang version and commit were not reported. The document pin is not a measured executable revision.
Measured checkpoint revision was not stated. The model-card revision is only a source and license reference.
Host RAM was not reported.
Duration, frames-per-second, and concurrency capacity were not independently reported. Duration is not derived from 107 frames.
20 NFE is not recorded here as an exact num_inference_steps schedule.
Component resident byte sizes were not measured.
No SSD path was measured.
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.
MiniMax-H3 model cardrevision 42ed227ee7df40d41602854ae760620d6eb651fe. Architecture and local-versus-hosted distinction. Not proof of the benchmark's checkpoint revision.
MiniMax H3 Community Licenserevision 42ed227ee7df40d41602854ae760620d6eb651fe. Dated 2026-08-02. Separate from runtime Apache-2.0.
SGLang Apache-2.0revision 2cd92d10b98a73de3952eafca915ae8954fcd375. Software license for the runtime document pin. 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.
This is one upstream single-GPU report. It is not a multi-host result and it has no four-GPU timing.
Rollback
Stop the attempted workload.
Restore the prior validated environment.
Recheck output and memory before another attempt.
No destructive deletion.
One 24GB RTX 4090 D finished an upstream SGLang T2VA run at 1344×768, 107 frames, and 20 NFE. The selected config is kitchen_int8 on DiT linear layers plus FlashAttention. Timed end-to-end time is 303.3 seconds and denoise time is 273.7 seconds, 1.34× the BF16 + FlashAttention baseline, with PSNR 24.81 dB against that baseline. Peak GPU stayed about 18 GB. This site did not reproduce the run, and the figure is not a calculator fit.
The document pin is SGLang 2cd92d10b98a73de3952eafca915ae8954fcd375, section rtx-4090-24-gb-single-gpu-run. The benchmark table sits around source line 2105. That pin is the cookbook text. The installed SGLang version and commit were not reported, so the document pin is not a measured executable revision. The measured checkpoint revision was not stated either. Model-card revision 42ed227ee7df40d41602854ae760620d6eb651fe is the license and architecture reference only.
What was selected
The workload uses the FL2VA checkpoint partition in T2VA mode. Sampler is euler. torch.compile and step caching are disabled. The DiT and the text encoder stream layerwise, with zero resident DiT layers and prefetch 1. The VAE stays on the device and uses tiled decode. The source warns that putting the VAE decoder into layerwise offload re-streams about 9 GiB on each of 167 decode tiles. That placement is not this selected report. The selected peak remains about 18 GB, in the source’s GB unit, for both the INT8 and BF16 streaming setups, because the peak is set by offload buffers and VAE decode rather than weight dtype.
kitchen_int8 quantizes DiT linears. It does not mean the text encoder or either VAE is INT8. Host RAM, duration, frames per second, and concurrency capacity were not independently reported. Do not turn 107 frames into a duration. Twenty NFE is not recorded here as an exact inference-step schedule. No SSD or NVMe path was measured.
The method is load, warmup with seed 0, then a timed pass with seed 42. Only the timed pass is reported.
Comparison rows, not the selected config
BF16 + FlashAttention took 405.6 seconds end-to-end and 370.2 seconds to denoise. kitchen_int8 with sol_attn took 223.9 / 203.9 seconds at 24.44 dB. kitchen_int8 with sage_attn took 174.9 / 154.2 seconds at 23.51 dB. The Sage-to-Sol hybrid took 163.8 / 143.1 seconds at 23.04 dB. Sol, Sage, and the hybrid change the attention algorithm. They are comparisons. The source default remains kitchen_int8 + FlashAttention.
Parts, license, and scope
The H3 text encoder uses Qwen3-VL-32B. Vision understanding belongs to that encoder. The omni transformer is a 33B dense model in the architecture description. This page does not add 33B to 32B and does not subtract the AdaLN branches from a checkpoint. Visual VAE and audio VAE are separate parts. Their resident byte sizes were not measured.
The model license is the MiniMax H3 Community License dated 2026-08-02. Applicable territory exclusions are the EU, the UK, the Republic of Korea, and the USA. Commercial products or services above US$20M yearly revenue need separate prior written authorization. Read the full license. This page does not infer your region or clear a deployment. SGLang’s Apache-2.0 license covers the runtime text, not the model. The encoder’s Apache-2.0 terms are separate.
Released local H3-Base is a 768p generator. Hosted Context-IR plus Regenerate-2K is a different service, and this site does not call it. Ref2VA is a different partition. The plain-text brief above is the planning copy. It contains no launch command.