Faster, Cheaper H3 Video on Blackhole®
ai& built its fastest, lowest-priced MiniMax H3 tier on Tenstorrent Galaxy®, and our benchmarks show it pulling ahead on heavier jobs.

MiniMax H3 is a 33-billion-parameter open-weight model that generates video and synchronized audio together in a single pass, for clips up to 15 seconds long.
In October, Japan-based inference company ai& published a detailed public study on serving H3 from its own data centers: “The road to faster, cheaper H3.” Its engineers ran H3 on NVIDIA B300, AMD MI355X, and Tenstorrent Blackhole Galaxy, tuned each platform, and launched a paid service based on the results. Here’s what stood out to us.
What ai& found
- Faster than real time. On four Galaxy systems, ai&'s H3 Fast tier (which uses a step-distilled version of the model) generates a 15-second clip in 11 seconds. A single system takes 19.
- Fastest full-model baseline. Without distillation, four Galaxy systems generated a 15-second clip in 68.7 seconds. Eight B300 GPUs took 102 seconds, and eight MI355X GPUs took 133.
- 4.4x lower price per second. H3 Fast runs on Blackhole Galaxy at $0.0125 per generated second, versus $0.055 for ai&’s full-model tier on GPUs. A 15-second clip costs under 19 cents.
- A quarter of the hardware cost. ai& estimates that a Galaxy system costs roughly a quarter of what it paid for its eight-GPU B300 system. A Galaxy system lists at $160,000.
- TT-DiT gets highlighted. ai&’s study points to our code as a concrete example of overlapping computation with data movement.
“Blackhole makes the economic opportunity tangible.”
— ai&, The road to faster, cheaper H3
Our own benchmarks of the full H3 model show Blackhole Galaxy pulling ahead as jobs get bigger. On a 5-second text-to-video clip, eight B300 GPUs are slightly faster than four Galaxy systems (18.6s vs. 21.9s). On a 15-second clip, four Galaxy systems are 1.5x faster (68.7s vs. 100.9s).
Reference-to-video jobs (Ref2VA), where the model also takes input images, video, and audio, save even more time. Four Galaxy systems beat eight B300 GPUs on six of seven full-model workloads at 15 seconds. On a 12-file reference set, they finish in 357 seconds instead of 614 (1.7x faster), saving over four minutes per clip.
| Workload (15s clip) | 4x Galaxy (s) | 8x B300 (s) | Time saved (s) |
|---|---|---|---|
| Text-to-video | 68.7 | 100.9 | 32.2 |
| First + last keyframe | 69.8 | 109.9 | 40.1 |
| Ref2VA: 3 images, 2 videos, 2 audio | 195.0 | 298.2 | 103.2 |
| Ref2VA: 6 images, 3 videos, 3 audio | 356.5 | 613.8 | 257.3 |
Full H3 model, no distillation. Tenstorrent benchmarks, mean of repeated runs. ai&’s study measured 102s for eight B300 GPUs on text-to-video.
ai& puts four Galaxy systems and eight B300 GPUs in the same purchase-price range, so the same budget produces more video on Blackhole Galaxy. A single Galaxy system, at roughly a quarter of the price, does even better: about 1.8x more video per dollar than eight B300 GPUs on text-to-video, and 2.7x on the heaviest Ref2VA job.
Blackhole Galaxy systems suit heavy jobs
Long clips and large reference sets give the model more to compute, and ai&’s profiling showed memory wasn’t the constraint: its MI355X runs used well under half the available memory.
At a similar purchase price, four Galaxy systems offer 92 PFLOPS of Block FP8 against 36 PFLOPS of dense FP8 for eight B300 GPUs (theoretical peaks, in different formats). On-chip SRAM, chip-to-chip Ethernet built into every Blackhole processor, and TT-DiT’s overlap of communication with computation keep data movement from eating into that compute. In ai&’s runs, going from one Galaxy system to four cut a 15-second clip from 229.1 to 68.7 seconds.
ai&’s GPU recipes (Cache-DiT plus FP8 on B300, and Cache-DiT plus MXFP6 attention on MI355X) cut full-model times by 2.4x and 3.3x. On Blackhole Galaxy, ai& went straight to step distillation. Cache-DiT and lower-precision attention haven’t been applied there yet, so the full-model results are only a starting point.
Read ai&’s full write-up, “The road to faster, cheaper H3,” and try H3 Fast on ai&’s Blackhole-based tier at aiand.com/video.


