mashbox-labs/Hy-Nomas-TPU

Hy-Nomas on Google Cloud TPU

Runs with mashbox run hy-nomas in the mashbox package, release 0.1.10.

Read our guide: run Hy-Nomas on TPU

Tencent's open video world model, playable in the browser from a TPU v6e-4. The full pipeline also fits on a single TPU v6e chip (32 GB), at 2.46 s of compute per second of video, main model quantized. Made by Tencent Hunyuan, ported to TPU by MASHBOX Labs.

• 1.045 s (median of 87 and of 27 windows on two chips): compute per second of video, quality preset, Tencent decoder, 1.00 = real time, on TPU v6e-4
• 0.689 s (median of 77 windows): compute per second of video, quality preset, MASHBOX Fast decoder, on TPU v6e-4
• 10 min 40.1 s: first start on a new chip without a matching compile cache, command to first frame on screen (weights downloaded and in the VM's memory; compiling included), on TPU v6e-4
• 1 min 14.8 s (median of 5 starts; 1 min 7.2 s to 1 min 43.6 s): later starts, command to first frame on screen (weights in the VM's memory), on TPU v6e-4
• 1 min 20.7 s: first start on a new chip with the compiled programs from this repo, command to first frame on screen (weights downloaded and in the VM's memory), on TPU v6e-4
• 2.46 s (60 s run): compute per second of video, full pipeline on a single TPU v6e chip (v6e-1, 32 GB), 720p, main model quantized; this version of mashbox run starts on a v6e-4

_{A world generated by Hy-Nomas on Google Cloud TPU. Shown smaller and at 8 frames per second; the model makes 1280x704 at 24.}

Quickstart

On a Google Cloud TPU VM (the setup guide gets you there from a new Google account; coding agents can follow AGENTS.md):

python3 -m pip install uv && python3 -m uv tool install --python 3.12 "mashbox[tpu,hy-nomas]" --extra-index-url https://download.pytorch.org/whl/cpu --index-strategy unsafe-best-match
~/.local/bin/mashbox run hy-nomas --ui

The app serves on the VM's 127.0.0.1:8790 only. Connect with gcloud compute tpus tpu-vm ssh hy-nomas --zone=asia-northeast1-b -- -L 8790:127.0.0.1:8790 (your chip's name and zone) and open http://127.0.0.1:8790 on your own machine. The start times below are with the weights downloaded and still in the VM's memory; the first start after the VM restarts reads them from disk again and takes longer.

Results

Chip	Measure	Result	Baseline
TPU v6e-4	Compute per second of video, quality preset, Tencent decoder, 1.00 = real time	1.045 s (median of 87 and of 27 windows on two chips)	1.00 s = real time
TPU v6e-4	Compute per second of video, quality preset, MASHBOX Fast decoder	0.689 s (median of 77 windows)	1.045 s with Tencent's decoder on the same chip type
TPU v6e-4	First start on a new chip without a matching compile cache, command to first frame on screen (weights downloaded and in the VM's memory; compiling included)	10 min 40.1 s	n/a
TPU v6e-4	Later starts, command to first frame on screen (weights in the VM's memory)	1 min 14.8 s (median of 5 starts; 1 min 7.2 s to 1 min 43.6 s)	n/a
TPU v6e-4	First start on a new chip with the compiled programs from this repo, command to first frame on screen (weights downloaded and in the VM's memory)	1 min 20.7 s	10 min 40.1 s without them (about 9 minutes of it compiling), same chip
TPU v6e-1 (one chip, 32 GB)	Compute per second of video, full pipeline on one chip, 720p, main model quantized	2.46 s (60 s run)	1.00 s = real time

Each chart names the hardware on both sides.

Files

This port loads the original weights unchanged from tencent/Hy-Nomas.

File	Format	Size (GB)	Bits per weight	Quality vs full precision	Speed
xla-cache/v6e4-jax0.11.2-libtpu0.0.48.1.tar	XLA compile cache (tar)	0.23	n/a	n/a (compiled programs, not weights)	first start on TPU v6e-4: first frame 1 min 20.7 s after the command, 10 min 40.1 s without it

Chips

Chip	Status	Note
TPU v5e	Not tested yet
TPU v6e	One chip	the full pipeline fits on a single TPU v6e chip (32 GB), 2.46 s of compute per second of video, main model quantized. This version of mashbox run starts on a v6e-4
TPU v5p slice (v5p-8)	Not tested yet
TPU v6e slice (v6e-4)	Supported	4 chips

Kernels

Compiled TPU kernels for this model only. mashbox run picks the file that matches your chip and JAX version and checks its sha256 against models.json. Without a match, the model runs on plain JAX, slower.

File	Chips	jax	jaxlib	libtpu
kernels/mashbox_hy_nomas_kernels-0.1.0-py3-none-any.whl	TPU v6e	0.11.2	0.11.2	0.0.48.*

Links

Code	mashbox-labs/mashbox
Guide	mashbox.gg/docs/models/hy-nomas
Original model	tencent/Hy-Nomas by Tencent Hunyuan
Setup from a new Google account	mashbox.gg/docs/google-cloud
Companies	hello@mashbox.gg

Licence

Weights: Apache-2.0 licence, from tencent/Hy-Nomas@4b8b4ed8; code: Apache-2.0 (MASHBOX Labs). LICENSE in this repo is Tencent Hunyuan's licence file from tencent/Hy-Nomas on Hugging Face, unchanged. Our own weights in this repo: decoders/hy-nomas-fast-decoder-v1.safetensors (the MASHBOX Fast Decoder): Apache-2.0. The compiled kernels in kernels/ are under their own licence: LICENSE-KERNELS. The compile cache in xla-cache/ is under the same licence.