Labs
Request access

MASHBOX Labs

MASHBOX Labs makes open models run faster and cheaper on TPUs and accelerated hardware, as well as exploring the future of the intersection of AI and gaming.

Install the mashbox command line

curl -fsSL https://mashbox.gg/install.sh | sh

macOS or Linux. On a Cloud TPU VM, the same line installs the TPU version.

Read the install script

Learn more

MASHBOX Steer

MASHBOX Steer: Qwen3.8-27B describes the sea in the Ancient dragon voice
Ancient dragon voice

See which neurons fire as a model reads your text, then turn them up or down and watch its replies change.

Try Steer

In MASHBOX for Mac

MASHBOX Steer: the feelings in play at each word of a chat, cold to hot
Feelings in play at each word

Models we run

38 models and files run in mashbox: 5 MASHBOX builds and 33 GGUF files of 20 models.

Hy-Nomas world model

Tencent Hunyuan's open video world model: text and movement keys in, video out, on a TPU v6e-4.

Our real-time research version

Hy-Nomas world model: an eagle flies over snowy peaks

GGUF files that run in our engine

GGUF files from other publishers, run as they are. Level and Runs on say how far each file has been checked; a speed shows only where we measured one on a chip. Each speed in these tables is from a short check of that one file. The Qwen3.8-27B numbers above are for our BQ8_R build.

0bserverx 1 fileQwen3.8-27B
ModelGGUF fileQuantLevelRuns onSpeed
Qwen3.8-27B
by Qwen
Q4_K_M
Quick check
Runs on TPU v6e-1 with the engine's TPU kernels: short check against llama.cpp (4 prompts x 32 tokens)
TPU v6e-1
39.28 tok/s, steady decode, 1 user, on TPU v6e-1
bartowski 7 filesFalcon3-1B-Instruct, Llama-3.2-1B-Instruct, Llama-3.2-3B-Instruct, Qwen2.5-0.5B-Instruct, Qwen2.5-1.5B-Instruct, Qwen2.5-7B-Instruct
ModelGGUF fileQuantLevelRuns onSpeed
Falcon3-1B-Instruct
by Technology Innovation Institute
Q8_0
Quick check
Runs on TPU v6e-1 with the engine's TPU kernels: short check against llama.cpp (4 prompts x 32 tokens)
TPU v6e-1
581.7 tok/s, steady decode, 1 user, on TPU v6e-1
Llama-3.2-1B-Instruct
by Meta
Q4_K_M
Quick check
Runs on TPU v6e-1 with the engine's TPU kernels: short check against llama.cpp (4 prompts x 32 tokens)
TPU v6e-1
472.41 tok/s, steady decode, 1 user, on TPU v6e-1
Llama-3.2-3B-Instruct
by Meta
IQ4_XS
Quick check
Runs on TPU v6e-1 with the engine's TPU kernels: short check against llama.cpp (4 prompts x 32 tokens)
TPU v6e-1
257.13 tok/s, steady decode, 1 user, on TPU v6e-1
Qwen2.5-0.5B-Instruct
by Qwen
Q4_K_M
Quick check
Runs on TPU v6e-1 with the engine's TPU kernels: short check against llama.cpp (4 prompts x 32 tokens)
TPU v6e-1
325.25 tok/s, steady decode, 1 user, on TPU v6e-1
Qwen2.5-1.5B-Instruct
by Qwen
Q4_K_M
Quick check
Runs on TPU v6e-1 with the engine's TPU kernels: short check against llama.cpp (4 prompts x 32 tokens)
TPU v6e-1
212.09 tok/s, steady decode, 1 user, on TPU v6e-1
Qwen2.5-1.5B-Instruct
by Qwen
Q5_K_M
Quick check
Runs on TPU v6e-1 with the engine's TPU kernels: short check against llama.cpp (4 prompts x 32 tokens)
TPU v6e-1
201.71 tok/s, steady decode, 1 user, on TPU v6e-1
Qwen2.5-7B-Instruct
by Qwen
Q4_K_M
Quick check
Runs on TPU v6e-1 with the engine's TPU kernels: short check against llama.cpp (4 prompts x 32 tokens)
TPU v6e-1
119.27 tok/s, steady decode, 1 user, on TPU v6e-1
byteshape 1 fileQwen3.8-27B
ModelGGUF fileQuantLevelRuns onSpeed
Qwen3.8-27B
by Qwen
IQ2_XXS
Quick check
Runs on TPU v6e-1 with the engine's TPU kernels: short check against llama.cpp (4 prompts x 32 tokens)
TPU v6e-1
11.78 tok/s, steady decode, 1 user, on TPU v6e-1
cdiamond 1 fileQwen3.8-27B
ModelGGUF fileQuantLevelRuns onSpeed
Qwen3.8-27B
by Qwen
MTP
Quick check
Runs on TPU v6e-1 with the engine's TPU kernels: short check against llama.cpp (4 prompts x 32 tokens)
TPU v6e-1
13.56 tok/s, steady decode, 1 user, on TPU v6e-1
DavidAU 2 filesQwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-MTP, Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU
ModelGGUF fileQuantLevelRuns onSpeed
Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-MTP
by DavidAU
Q4_K_M
Quick check
Runs on TPU v6e-1 with the engine's TPU kernels: short check against llama.cpp (4 prompts x 32 tokens)
TPU v6e-1
108.8 tok/s, steady decode, 1 user, on TPU v6e-1
Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU
by DavidAU
Q4_K_M
Quick check
Runs on TPU v6e-1 with the engine's TPU kernels: short check against llama.cpp (4 prompts x 32 tokens)
TPU v6e-1
39.21 tok/s, steady decode, 1 user, on TPU v6e-1
ggml-org 1 fileQwen3.8-27B
ModelGGUF fileQuantLevelRuns onSpeed
Qwen3.8-27B
by Qwen
Q4_K_M
Quick check
Runs on TPU v6e-1 with the engine's TPU kernels: short check against llama.cpp (4 prompts x 32 tokens)
TPU v6e-1
41.48 tok/s, steady decode, 1 user, on TPU v6e-1
HauhauCS 1 fileQwen3.8-27B
ModelGGUF fileQuantLevelRuns onSpeed
Qwen3.8-27B
by Qwen
Q4_K_P
Quick check
Runs on TPU v6e-1 with the engine's TPU kernels: short check against llama.cpp (4 prompts x 32 tokens)
TPU v6e-1
35.5 tok/s, steady decode, 1 user, on TPU v6e-1
huihui-ai 1 fileQwen3.8-27B
ModelGGUF fileQuantLevelRuns onSpeed
Qwen3.8-27B
by Qwen
Q4_K_M
Quick check
Runs on TPU v6e-1 with the engine's TPU kernels: short check against llama.cpp (4 prompts x 32 tokens)
TPU v6e-1
32.82 tok/s, steady decode, 1 user, on TPU v6e-1
ibm-granite 1 filegranite-3.3-2b-instruct
ModelGGUF fileQuantLevelRuns onSpeed
granite-3.3-2b-instruct
by IBM
Q4_K_M
Quick check
Runs on TPU v6e-1 with the engine's TPU kernels: short check against llama.cpp (4 prompts x 32 tokens)
TPU v6e-1
238.1 tok/s, steady decode, 1 user, on TPU v6e-1
JonathanColetti 1 fileQwen3.8-27B
ModelGGUF fileQuantLevelRuns onSpeed
Qwen3.8-27B
by Qwen
Q4_K_M
Quick check
Runs on TPU v6e-1 with the engine's TPU kernels: short check against llama.cpp (4 prompts x 32 tokens)
TPU v6e-1
39.12 tok/s, steady decode, 1 user, on TPU v6e-1
LGAI-EXAONE 1 fileEXAONE-3.5-2.4B-Instruct
ModelGGUF fileQuantLevelRuns onSpeed
EXAONE-3.5-2.4B-Instruct
by LG AI Research
Q4_K_M
Quick check
Runs on TPU v6e-1 with the engine's TPU kernels: short check against llama.cpp (4 prompts x 32 tokens)
TPU v6e-1
297.52 tok/s, steady decode, 1 user, on TPU v6e-1
lmstudio-community 2 filesQwen3.5-9B, Qwen3.8-27B
ModelGGUF fileQuantLevelRuns onSpeed
Qwen3.5-9B
by Qwen
Q4_K_M
Quick check
Runs on TPU v6e-1 with the engine's TPU kernels: short check against llama.cpp (4 prompts x 32 tokens)
TPU v6e-1
111.73 tok/s, steady decode, 1 user, on TPU v6e-1
Qwen3.8-27B
by Qwen
Q4_K_M
Quick check
Runs on TPU v6e-1 with the engine's TPU kernels: short check against llama.cpp (4 prompts x 32 tokens)
TPU v6e-1
39.08 tok/s, steady decode, 1 user, on TPU v6e-1
peculiar-ragdoll 1 fileQwen3.8-27B
ModelGGUF fileQuantLevelRuns onSpeed
Qwen3.8-27B
by Qwen
Q4_K_M
Quick check
Runs on TPU v6e-1 with the engine's TPU kernels: short check against llama.cpp (4 prompts x 32 tokens)
TPU v6e-1
32.08 tok/s, steady decode, 1 user, on TPU v6e-1
Qwen 5 filesQwen2.5-0.5B-Instruct, Qwen2.5-1.5B-Instruct, Qwen2.5-Coder-1.5B-Instruct, Qwen3-14B, Qwen3-4B
ModelGGUF fileQuantLevelRuns onSpeed
Qwen2.5-0.5B-Instruct
by Qwen
q4_k_m
Quick check
Runs on TPU v6e-1 with the engine's TPU kernels: short check against llama.cpp (4 prompts x 32 tokens)
TPU v6e-1
326.75 tok/s, steady decode, 1 user, on TPU v6e-1
Qwen2.5-1.5B-Instruct
by Qwen
q4_k_m
Quick check
Runs on TPU v6e-1 with the engine's TPU kernels: short check against llama.cpp (4 prompts x 32 tokens)
TPU v6e-1
211.71 tok/s, steady decode, 1 user, on TPU v6e-1
Qwen2.5-Coder-1.5B-Instruct
by Qwen
q4_k_m
Quick check
Runs on TPU v6e-1 with the engine's TPU kernels: short check against llama.cpp (4 prompts x 32 tokens)
TPU v6e-1
210.52 tok/s, steady decode, 1 user, on TPU v6e-1
Qwen3-14B
by Qwen
Q4_K_M
Quick check
Runs on TPU v6e-1 with the engine's TPU kernels: short check against llama.cpp (4 prompts x 32 tokens)
TPU v6e-1
63.7 tok/s, steady decode, 1 user, on TPU v6e-1
Qwen3-4B
by Qwen
Q4_K_M
Quick check
Runs on TPU v6e-1 with the engine's TPU kernels: short check against llama.cpp (4 prompts x 32 tokens)
TPU v6e-1
166.3 tok/s, steady decode, 1 user, on TPU v6e-1
unsloth 7 filesgemma-3-1b-it, Qwen3-0.6B, Qwen3-1.7B, Qwen3-4B-Instruct-2507, Qwen3.5-4B, Qwen3.5-9B, Qwen3.8-27B
ModelGGUF fileQuantLevelRuns onSpeed
gemma-3-1b-it
by Google DeepMind
Q4_K_M
Quick check
Runs on TPU v6e-1 with the engine's TPU kernels: short check against llama.cpp (4 prompts x 32 tokens)
TPU v6e-1
212.82 tok/s, steady decode, 1 user, on TPU v6e-1
Qwen3-0.6B
by Qwen
Q8_0
Quick check
Runs on TPU v6e-1 with the engine's TPU kernels: short check against llama.cpp (4 prompts x 32 tokens)
TPU v6e-1
416.18 tok/s, steady decode, 1 user, on TPU v6e-1
Qwen3-1.7B
by Qwen
Q4_0
Quick check
Runs on TPU v6e-1 with the engine's TPU kernels: short check against llama.cpp (4 prompts x 32 tokens)
TPU v6e-1
239.61 tok/s, steady decode, 1 user, on TPU v6e-1
Qwen3-4B-Instruct-2507
by Qwen
Q4_K_M
Quick check
Runs on TPU v6e-1 with the engine's TPU kernels: short check against llama.cpp (4 prompts x 32 tokens)
TPU v6e-1
166.67 tok/s, steady decode, 1 user, on TPU v6e-1
Qwen3.5-4B
by Qwen
Q4_K_M
Quick check
Runs on TPU v6e-1 with the engine's TPU kernels: short check against llama.cpp (4 prompts x 32 tokens)
TPU v6e-1
171.83 tok/s, steady decode, 1 user, on TPU v6e-1
Qwen3.5-9B
by Qwen
Q4_K_M
Quick check
Runs on TPU v6e-1 with the engine's TPU kernels: short check against llama.cpp (4 prompts x 32 tokens)
TPU v6e-1
112.04 tok/s, steady decode, 1 user, on TPU v6e-1
Qwen3.8-27B
by Qwen
Q4_K_M
Quick check
Runs on TPU v6e-1 with the engine's TPU kernels: short check against llama.cpp (4 prompts x 32 tokens)
TPU v6e-1
32.36 tok/s, steady decode, 1 user, on TPU v6e-1

Every file with its run command

Track2Rig Experimental

Track2Rig: a pelican spider's skeleton and bone motion, recovered from its 3D point tracks

3D point tracks in; skeleton, skin and bone motion out as GLB. Error on our test set, not on this clip: 0.784%, against 0.915% for SSDR.

See Track2Rig

Start here

From a new Google account to an open model answering on your own TPU, then to your coding agent.

MASHBOX Labs onMASHBOX Labs on

Latest release: mashbox 0.1.15. Open source

  1. Install mashbox

    curl -fsSL https://mashbox.gg/install.sh | sh

    On your Mac or Linux machine, without sudo.

    Read the install script

  2. Set up Google Cloud

    Make a Google Cloud project with billing and the TPU API turned on.

    Run on Google Cloud TPU, steps 1 to 5

  3. Bring up a TPU and run a model

    mashbox tpu up qwen3-8-27b

    It shows the price and asks before it creates anything.

  4. Connect your coding agent

    mashbox tpu up qwen3-8-27b --agents --restart

    Claude Code, Codex and Pi can use the model you serve as their LLM.

    Use with coding agents, from step 2

  5. Read the docs

    The FAQ, troubleshooting, which TPU to pick, the licences, and an llms.txt your coding agent can read.

    Docs

Earlier results

Every number is our own measurement unless it names a source. H100 speeds are published figures: EdenLive's tech report by VAST, section 5.4, and llama.cpp's scoreboard.

All results, with sources, on Research

AI and gaming

3D creatures from a text prompt, motion for any creature's skeleton and voices from a plain description, running on TPU, plus rigs recovered from 3D point tracks.

MASHBOX Motion AnyMASHBOX Voice

Rigged, animated 3D creatures

Live demo

Up to 200 characters of prompt in. One GLB out, rigged and textured. Each creature comes with its own motion clips. Pick one and drag to turn the model.

A 3D eagle

  • 305 s from click to downloads for a full text-to-rigged creature in the live demo.
  • Chip cost per creature for the first job on a new TPU v6e chip, with the models loading: $0.065 at the Google Cloud spot list price, $0.318 at the on-demand list price.

Styles: realistic (default), anime, cartoon, storybook, horror. POST /v1/creatures.

Eagle, made from a text prompt, with motion from our motion model.
Ghoul, made from a text prompt, with motion from our motion model.
Tarantula, made from a text prompt: textured and rigged.
Stag, made from a text prompt, with motion from our motion model.
Pegasus, made from a text prompt: textured and rigged.
Border collie, made from a text prompt: textured and rigged.

For businesses

What we sell

Lamproite is our software layer for AI on TPUs. One platform, eight products.

Lamproite PacksReady-to-run models. One model or pipeline, proven on a real TPU chip and frozen, with a self-test and a cost card.
ScaifeServing engine. 2.10 times vLLM TPU 0.31 on Llama 3.2 1B at 256 conversations.
BSHJWeight format. Bit-exact weights, built for fast cold starts.
CuletKernel library. Each kernel has a reference, a test and a provenance record.
Culet 3DOur replacements for CUDA-only 3D libraries.
MasterstoneProof report. A model graded against its own reference: parity, speed and cost per output on your chip.
RecutPorting. Your model on TPU, delivered as a Lamproite Pack with a Masterstone report.
Lamproite ManagedManaged serving. Your brand, our chips: a private endpoint per Pack.

A Pack is one catalog model ported to Lamproite, our TPU software platform. You run it on your own Google Cloud TPUs, under licence and NDA. Recut is porting to order. Send an open model and reference outputs from your current setup, and the port comes back with a Masterstone report graded against them.

Compiled software or a managed endpoint, under licence and NDA. Write to [email protected].

All modelsGallery

Models and code used in these resultsHow we measure

Install the mashbox command line

curl -fsSL https://mashbox.gg/install.sh | sh

macOS or Linux. On a Cloud TPU VM, the same line installs the TPU version.

Read the install script