MASHBOXLabs
Request access

Results from MASHBOX Labs

Updated 24 October 2026.

Every number is our own measurement unless it names a source.

Results

Live video23 October 2026

EdenLive 5B streams 87.4 frames a second at 832x480 on one TPU v6e chip

As of 23 October 2026: the same model, EdenLive 5B by VAST, built on Wan2.2, streaming at 832×480: 87.4 frames a second on one Google TPU v6e chip in our engine (87.38 and 87.67 in two runs), 1.19 times the 73.52 frames a second on one NVIDIA H100 in EdenLive’s published figure.

Frames a second at 832x480, higher is better EdenLive on one H100, publishedOurs, one TPU v6e

832x480
73.5287.4

H100: EdenLive tech report by VAST, section 5.4, Tables 2 and 3.

Live video22 October 2026

EdenLive 5B by VAST (built on Wan2.2) runs live at 1280x704 and 39.4 frames a second in our engine, with the first chunk in 256 ms. EdenLive's own report gives 29.8 frames a second on an NVIDIA H100. Measured 22 October 2026.

Frames a second at 1280x704, higher is better EdenLive on one H100, publishedOurs, one TPU v6e

1280x704
29.839.4

H100: EdenLive tech report by VAST, section 5.4, Tables 2 and 3.

Video21 October 2026

A final-quality 81-frame Wan 2.2 A14B clip at 832x464 in 16 seconds. As of 21 October 2026.

Serving engine19 October 2026

Our serving engine runs Llama 3.2 1B at 2.10 times the throughput of vLLM-TPU 0.31

On Llama 3.2 1B with the same weights: 47,895 tokens a second at 256 conversations, 2.10 times Google's vLLM-TPU 0.31 (22,826). Measured 19 October 2026.

Llama 3.2 1B, same chip and weights, 256 conversations, tokens a second, higher is better

vLLM-TPU 0.31
22,826
Ours
47,895

Language models19 October 2026

Qwen3-30B-A3B at 4-bit: 598.7 tokens a second for one conversation, 2.10 times the published RTX 5090 llama.cpp result (285.5). Measured 19 October 2026.

Qwen3-30B-A3B at 4-bit, tokens a second for one conversation, higher is better

RTX 5090, llama.cpp, published
285.5
Ours
598.7

Language models22 October 2026

Xiaomi MiMo-V2.6-Flash against the fastest listed provider

Xiaomi MiMo-V2.6-Flash: 180.4 tokens a second for one conversation, 1.22 times the fastest provider listed on Artificial Analysis (Makora, 147.8), with a 10,000-token prompt and 5.31 s to first token. Measured 22 October 2026.

Tokens a second for one conversation, higher is better

Makora
147.8
Ours
180.4

llama.cpp on TPU19 October 2026

Llama 2 7B at 4-bit through llama.cpp on TPU: 249.9 tokens a second, above the RTX 4090 (188.96) and A100 80GB (200.90) rows of llama.cpp's scoreboard. Measured 19 October 2026.

Llama 2 7B at 4-bit, tokens a second, llama.cpp scoreboard, higher is better

RTX 4090
188.96
A100 80GB
200.90
Ours, llama.cpp on TPU
249.9

Rigging22 October 2026

Track2Rig recovers skeleton, skin and motion with 0.678% error, against 0.895% for SSDR

From noisy 3D point tracks of 16 creatures it never saw (32 walk and idle clips), Track2Rig recovers a skeleton, skin and motion with a reconstruction error of 0.678% of body size on held-out surface points, against 0.895% for SSDR, the standard skinning-decomposition method, at the same bone count. Lower on all 32 clips. Pre-registered and read once, 22 October 2026.

Error on held-out surface points, % of body size, lower is better

SSDR
0.895
Track2Rig
0.678

See the clips

Voice18 and 19 October 2026

A spoken NPC reply starts 1.39 s after the player stops talking

A full spoken NPC reply starts 1.39 s (median) after the player stops talking, speech recognition and network included, on par with Convai's published 1.42 s in Unreal Engine 5. Measured 18 and 19 October 2026.

Seconds from the end of the player's speech to the first reply audio, lower is better

Convai in Unreal Engine 5, published
1.42
Ours, median
1.39

Convai: Frontiers in Computer Science, 2026, Table 4.

Porting21 to 23 October 2026

NVIDIA-only code replaced with our own, and checked against the original output

142 of 142

flash-attn and xformers

flash-attn and xformers, CUDA-only attention libraries, run the sparse transformer stages of Microsoft's TRELLIS.2 (AniGen and Pixal3D use them too). We replaced them. On one TPU v6e chip, those TRELLIS.2 stages matched the original model in 142 of 142 stage-by-stage checks. Measured 23 October 2026.

85 of 88

FlexGEMM

FlexGEMM, a CUDA-only sparse 3D convolution library, sits inside Microsoft's TRELLIS.2 image-to-3D model (its shape and texture decoders) and Pixal3D. We replaced it with our own code. On one TPU v6e chip, TRELLIS.2's sparse decoder stages matched the original model in 85 of 88 stage-by-stage checks. Measured 23 October 2026.

72 of 72

Audio2Face-3D without TensorRT

NVIDIA's SDK runs its Audio2Face-3D v3.0 model on NVIDIA's TensorRT. We run the network natively on a TPU v5e chip, and it matched the reference output in 72 of 72 test cases (worst relative error 1.44e-4). Measured 23 October 2026.

12 of 12

gsplat

gsplat is the CUDA-only renderer used to draw 3D Gaussian-splat scenes. Our own TPU renderer reproduced its output on 12 of 12 held-out views of a room scene, with a lowest PSNR of 103.9 dB against gsplat 1.5.3, on a TPU v5p. Measured 21 October 2026.

Cost per output

Cost next to an H100

Cost per output, with both chips at Google Cloud on-demand list prices: TPU v6e at $2.70 and H100 at $11.06 per chip-hour. Our speeds are measured; each H100 speed is the published figure cited below the table.

OutputOurs, one TPU v6eH100, citedOur costH100 costOurs vs H100
Live video, EdenLive 5B, 1280x704, per 1,000 frames39.4 fps29.8 fps$0.019$0.1030.18x
Live video, EdenLive 5B, 832x480, per 1,000 frames87.4 fps73.52 fps$0.0086$0.0420.21x
Llama 2 7B at 4-bit, one user, per 1M tokens249.9 tok/s280.74 tok/s$3.00$10.940.27x

Ours: measured on one chip, benchmark: row 1 on 22 October, row 2 as of 23 October, row 3 on 19 October 2026. H100 reference: cited. Rows 1 and 2: EdenLive's own tech report by VAST (section 5.4, Tables 2 and 3). Row 3: llama.cpp's CUDA scoreboard for Llama 2 7B, H100 80GB row, read 22 October 2026. List prices read 19 October 2026.

Our cost next to the closest product

OutputOur costClosest product, list price
MASHBOX Creature, per creature$0.136$0.19-0.21, Ludo
Track2Rig, per rig$0.000119$0.03, Ludo
MASHBOX Stride, per motion clip$0.000316$0.06, fal

Our cost: run time we measured at today's small scale, priced at Google Cloud on-demand list rates, 23 October 2026; Creature with the pipeline already loaded; Stride uses the higher of our two cost measurements (24 October 2026). List prices read 23 October 2026. Nothing on the market takes the same input as Track2Rig, Stride or Creature, so those rows use the closest product: Ludo for Track2Rig and Creature (Ludo's creature motions are each under a second long), a different motion model on fal for Stride.

Models

13 models and pipelines in the catalog.

Browse models

Models and code used in these results

  • Z-Image Turbo, Tongyi-MAI (Alibaba)
  • Wan2.2, Wan team (Alibaba)
  • Qwen3-30B-A3B, Qwen team (Alibaba)
  • Llama 3.1 8B Instruct, Llama 3.2 1B and Llama 2 7B, Meta
  • EdenLive 5B, VAST
  • FLUX.2 [klein] 4B, Black Forest Labs
  • TRELLIS.2, Microsoft
  • TokenRig, TripoSG and AniGen, VAST
  • Audio2Face-3D, NVIDIA
  • Harrier 27B, Microsoft
  • Granite Speech 4.1 2B, IBM
  • Z-Image ControlNet Union, Alibaba PAI
  • InfiniteTalk, MeiGen-AI
  • Wan 2.2 Animate, Wan team (Alibaba)
  • Gemma 4, Google
  • Pixal3D, Tencent ARC
  • MiniCPM-o 4.5, OpenBMB
  • MiMo-V2.6-Flash, Xiaomi
  • Maya1, Maya Research
  • llama.cpp
  • vLLM TPU
  • MaxDiffusion

About

MASHBOX Labs is a research lab and software layer for AI on TPUs. Write to [email protected] or find us on X @mashboxgg.