All models
Every model mashbox runs on Google Cloud TPU, with the chip it needs and one measured number each. Each name opens its guide.
| Model | What it does | Chips | Measured | Weights |
|---|---|---|---|---|
| Hy-Nomas by Tencent Hunyuan |
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. | TPU v6e (v6e-1): 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-4TPU v5e: Not tested yet TPU v5p slice (v5p-8): Not tested yet TPU v6e slice (v6e-4): Supported |
1.045 s compute per second of video, quality preset, Tencent decoder, 1.00 = real time (TPU v6e-4) | Hy-Nomas-TPU |
| Qwen3.8-27B by Qwen |
Qwen’s 27B hybrid-attention language model on one Google Cloud TPU v6e chip, with an OpenAI-compatible server. | TPU v6e (v6e-1): Supported TPU v5e (v5e-4): Supported TPU v5p slice (v5p-8): Not tested yet TPU v6e slice (v6e-4): Supported |
38.1 vs 31.4 tok/s vs vLLM on the same chip, 1 request at a time (TPU v6e-1) | Qwen3.8-27B-TPU |
| Track2Rig by MASHBOX Labs |
Our own model: dense 3D point tracks of a moving creature in, a rig out (bones, skin weights, per-frame bone transforms and a skeleton). Runs on a CPU. It does not work well on many clips yet. Newer weights will be published as dated versions. | TPU v6e: Supported TPU v5e: Not tested yet TPU v5p slice (v5p-8): Not tested yet TPU v6e slice (v6e-4): Not tested yet CPU: Supported |
0.784% vs 0.915% held-out error vs SSDR at the same bone count, synthetic test set with noisy, partly hidden tracks, % of body size, lower is better (CPU (AMD EPYC 7502P)) | Track2Rig |
| TRELLIS.2 by Microsoft |
Microsoft’s open image-to-3D model: one RGBA picture in, a textured GLB mesh out, on a TPU v6e-1 VM (the model on the chip, the GLB export on the VM’s CPU). | TPU v5e: Not tested yet TPU v6e (v6e-1): Supported TPU v5p slice (v5p-8): Not tested yet TPU v6e slice (v6e-4): Not tested yet |
15 min 32.4 s to 30 min 11.7 s one picture in, a textured GLB out: the whole command (TPU v6e-1) | TRELLIS.2-TPU |
These are the MASHBOX builds. The GGUF files that run in our engine, and the count of everything in this version, are on Supported models.
Which chip
Section titled “Which chip”| Model | Chip | Memory on the chip | What the model needs | Download | Price, on demand | Zones that granted a chip |
|---|---|---|---|---|---|---|
| Qwen3.8-27B | TPU v6e-1 | 32 GB (31.24 GiB) | 25.09 GiB in use with 8 request slots | 26.94 GB file, 8.01 bits per weight | $2.70 an hour in US regions | europe-west4-a, us-east5-b |
| Qwen3.8-27B | TPU v6e-4 | 4 x 32 GB | 4 chips, split across them automatically | 26.94 GB file, 8.01 bits per weight | $10.80 an hour in US regions | |
| Qwen3.8-27B | TPU v5e-4 | 4 x 16 GB | 4 chips, split across them automatically (the weights alone are about 27 GB, so one 16 GB v5e chip is too small) | 26.94 GB file, 8.01 bits per weight | $4.80 an hour in US regions | |
| Hy-Nomas | TPU v6e-4 | 4 x 32 GB | 4 chips: the main model on 4, the video decoder on 3 | 87.4 GB (Tencent’s weights and the text encoder) | $10.80 an hour in US regions | us-east5-b, asia-northeast1-b, europe-west4-a (queued requests) |
| Hy-Nomas | TPU v6e-1 | 32 GB | the full pipeline fits on one chip: 2.46 s of compute per second of video, main model quantized (mashbox run starts on a v6e-4) |
$2.70 an hour in US regions | ||
| Hy-Nomas | TPU v5e | 16 GB | Not tested yet |
Prices are Google’s on-demand list prices; a chip bills from create to delete. Memory in GiB is what JAX reported on the chip.