Supported models
Every model and file in this version of mashbox’s catalog, on one page in two parts. MASHBOX builds are the models we port or make ourselves, each with its measured number against its baseline. GGUF files that run in our engine are files from other publishers that mashbox runs as they are, one section per model family, with each file’s maker, architecture and quant. The Level column says how far each one has been checked. mashbox models prints the same rows from the package’s own catalog, and mashbox models --json prints them as JSON.
mashbox modelsEach row names the chips it runs on and the command that runs it. A GGUF link opens the exact file revision that ran.
37 entries: 3 MASHBOX ports, 1 MASHBOX model and 33 GGUF files of 20 models.
MASHBOX builds
Section titled “MASHBOX builds”Models we port or make ourselves. A BSHJ build is our weights file for the model, run on our TPU kernels. Measured is each one’s number against its baseline, on the chip named.
| Model | Format | Level | Runs on | Measured | Run |
|---|---|---|---|---|---|
| Hy-Nomas Made by Tencent Hunyuan, ported to TPU by MASHBOX Labs |
TPU port World model: text and movement keys in, video out Hugging Face |
MASHBOX build | TPU v6e-4 | Compute per second of video, quality preset, Tencent decoder, 1.00 = real time: 1.045 s (TPU v6e-4) Compute per second of video, quality preset, MASHBOX Fast decoder: 0.689 s (TPU v6e-4) |
~/.local/bin/mashbox run hy-nomas --ui |
| Qwen3.8-27B Made by Qwen, ported to TPU by MASHBOX Labs |
BSHJ build: BQ8_R LLM: text in, text out Hugging Face |
MASHBOX build | TPU v5e-4, TPU v6e-1, TPU v6e-4 | 38.1 vs 31.4 tok/s vs vLLM on the same chip, 1 request at a time (TPU v6e-1) 74.84 tok/s with speculative decoding (–mtp 2), 1 request at a time, greedy; 37.77 tok/s with it off (TPU v6e-1) 326.6 vs 1,606.7 s first start to ready, weights on disk, empty compile caches, vs vLLM on the same chip type (TPU v6e-1) |
~/.local/bin/mashbox run qwen3-8-27b |
| Track2Rig (Experimental) Trained by MASHBOX Labs |
MASHBOX model Rigging: 3D point tracks in, rig (JSON and glTF) out Hugging Face |
MASHBOX build | CPU, TPU v6e | 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) 6.24% vs 7.73% joint error vs SSDR, % of body size, lower is better (CPU, AMD EPYC 7502P) |
mashbox run track2rig examples/track2rig/tracks/grison_walk.npz -o grison_rig |
| TRELLIS.2 (Experimental) Made by Microsoft, ported to TPU by MASHBOX Labs |
TPU port image-to-3d: an RGBA image (background cut out) in, a textured 3D mesh (GLB) out Hugging Face |
MASHBOX build | TPU v6e-1 | One picture in, a textured GLB out: the whole command: 15 min 32.4 s to 30 min 11.7 s (TPU v6e-1) Picture in to mesh out: the model’s stages on the chip, with the host work between them: 46.7 s to 1 min 37.5 s (TPU v6e-1) |
~/.local/bin/mashbox run trellis-2 --image object.png --out object.glb |
GGUF files that run in our engine
Section titled “GGUF files that run in our engine”GGUF files from other publishers, run as they are by mashbox. Level and Runs on say how far each file has been checked; a speed shows only where a chip run measured it on our kernels.
| Model | GGUF file | Architecture | Quant | Level | Runs on | Speed | Run |
|---|---|---|---|---|---|---|---|
| Qwen2.5-0.5B-Instruct by Qwen |
Qwen2.5- from bartowski, 0.4 GB |
qwen2 |
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 | 325.35 tok/s, steady decode, 1 user, on TPU v6e-1 | mashbox run bartowski/Qwen2.5-0.5B-Instruct-GGUF:Q4_K_M |
| Qwen2.5-0.5B-Instruct by Qwen |
qwen2.5- from Qwen, 0.49 GB |
qwen2 |
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 | 322.16 tok/s, steady decode, 1 user, on TPU v6e-1 | mashbox run Qwen/Qwen2.5-0.5B-Instruct-GGUF:q4_k_m |
| Qwen2.5-1.5B-Instruct by Qwen |
Qwen2.5- from bartowski, 0.99 GB |
qwen2 |
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 | 209.35 tok/s, steady decode, 1 user, on TPU v6e-1 | mashbox run bartowski/Qwen2.5-1.5B-Instruct-GGUF:Q4_K_M |
| Qwen2.5-1.5B-Instruct by Qwen |
qwen2.5- from Qwen, 1.12 GB |
qwen2 |
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 | 210.41 tok/s, steady decode, 1 user, on TPU v6e-1 | mashbox run Qwen/Qwen2.5-1.5B-Instruct-GGUF:q4_k_m |
| Qwen2.5-1.5B-Instruct by Qwen |
Qwen2.5- from bartowski, 1.13 GB |
qwen2 |
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 | 200.56 tok/s, steady decode, 1 user, on TPU v6e-1 | mashbox run bartowski/Qwen2.5-1.5B-Instruct-GGUF:Q5_K_M |
| Qwen2.5-7B-Instruct by Qwen |
Qwen2.5- from bartowski, 4.68 GB |
qwen2 |
Q4_K_M | Quick check Runs on one chip of a TPU v6e-4 with the engine’s TPU kernels: short check against llama.cpp (4 prompts x 32 tokens) |
one chip of a TPU v6e-4 | 116.93 tok/s, steady decode, 1 user, on one chip of a TPU v6e-4 | mashbox run bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M |
| Qwen2.5-Coder-1.5B-Instruct by Qwen |
qwen2.5- from Qwen, 1.12 GB |
qwen2 |
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 | 209.77 tok/s, steady decode, 1 user, on TPU v6e-1 | mashbox run Qwen/Qwen2.5-Coder-1.5B-Instruct-GGUF:q4_k_m |
| Qwen3-0.6B by Qwen |
Qwen3- from unsloth, 0.64 GB |
qwen3 |
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 | 415.85 tok/s, steady decode, 1 user, on TPU v6e-1 | mashbox run unsloth/Qwen3-0.6B-GGUF:Q8_0 |
| Qwen3-1.7B by Qwen |
Qwen3- from unsloth, 1.06 GB |
qwen3 |
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 | 239.5 tok/s, steady decode, 1 user, on TPU v6e-1 | mashbox run unsloth/Qwen3-1.7B-GGUF:Q4_0 |
| Qwen3-14B by Qwen |
Qwen3- from Qwen, 9 GB |
qwen3 |
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 | 61.69 tok/s, steady decode, 1 user, on TPU v6e-1 | mashbox run Qwen/Qwen3-14B-GGUF:Q4_K_M |
| Qwen3-4B by Qwen |
Qwen3- from Qwen, 2.5 GB |
qwen3 |
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 | 163.86 tok/s, steady decode, 1 user, on TPU v6e-1 | mashbox run Qwen/Qwen3-4B-GGUF:Q4_K_M |
| Qwen3-4B-Instruct-2507 by Qwen |
Qwen3- from unsloth, 2.5 GB |
qwen3 |
Q4_K_M | Quick check Runs on one chip of a TPU v6e-4 with the engine’s TPU kernels: short check against llama.cpp (4 prompts x 32 tokens) |
one chip of a TPU v6e-4 | 163.83 tok/s, steady decode, 1 user, on one chip of a TPU v6e-4 | mashbox run unsloth/Qwen3-4B-Instruct-2507-GGUF:Q4_K_M |
| Qwen3.5-4B by Qwen |
Qwen3.5- from unsloth, 2.74 GB |
qwen35 |
Q4_K_M | Quick check Runs on one chip of a TPU v6e-4 with the engine’s TPU kernels: short check against llama.cpp (4 prompts x 32 tokens) |
one chip of a TPU v6e-4 | 168.75 tok/s, steady decode, 1 user, on one chip of a TPU v6e-4 | mashbox run unsloth/Qwen3.5-4B-GGUF:Q4_K_M |
| Qwen3.5-9B by Qwen |
Qwen3.5- from lmstudio-community, 5.63 GB |
qwen35 |
Q4_K_M | Quick check Runs on one chip of a TPU v6e-4 with the engine’s TPU kernels: short check against llama.cpp (4 prompts x 32 tokens) |
one chip of a TPU v6e-4 | 109.09 tok/s, steady decode, 1 user, on one chip of a TPU v6e-4 | mashbox run lmstudio-community/Qwen3.5-9B-GGUF:Q4_K_M |
| Qwen3.5-9B by Qwen |
Qwen3.5- from unsloth, 5.68 GB |
qwen35 |
Q4_K_M | Quick check Runs on one chip of a TPU v6e-4 with the engine’s TPU kernels: short check against llama.cpp (4 prompts x 32 tokens) |
one chip of a TPU v6e-4 | 109.6 tok/s, steady decode, 1 user, on one chip of a TPU v6e-4 | mashbox run unsloth/Qwen3.5-9B-GGUF:Q4_K_M |
| Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-MTP by DavidAU |
Qwen3.5- from DavidAU, 6.98 GB |
qwen35 |
Q4_K_M | Quick check Runs on one chip of a TPU v6e-4 with the engine’s TPU kernels: short check against llama.cpp (4 prompts x 32 tokens) |
one chip of a TPU v6e-4 | 108.05 tok/s, steady decode, 1 user, on one chip of a TPU v6e-4 | mashbox run DavidAU/Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-MTP-GGUF:Q4_K_M |
| Qwen3.8-27B by Qwen |
Qwen3.8- from byteshape, 8.84 GB |
qwen35 |
IQ2_XXS | Quick check Runs on one chip of a TPU v6e-4 with the engine’s TPU kernels: short check against llama.cpp (4 prompts x 32 tokens) |
one chip of a TPU v6e-4 | 11.8 tok/s, steady decode, 1 user, on one chip of a TPU v6e-4 | mashbox run byteshape/Qwen3.8-27B-GGUF:IQ2_XXS |
| Qwen3.8-27B by Qwen |
Qwen3.8- from cdiamond, 17.13 GB |
qwen35 |
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 | 13.53 tok/s, steady decode, 1 user, on TPU v6e-1 | mashbox run cdiamond/Qwen3.8-27B-iMatrix-NVFP4-MTP-GGUF:MTP |
| Qwen3.8-27B by Qwen |
Dirk- from peculiar-ragdoll, 16.46 GB |
qwen35 |
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 | 32.28 tok/s, steady decode, 1 user, on TPU v6e-1 | mashbox run peculiar-ragdoll/Dirk-Qwen3.8-27B-GGUF:Q4_K_M |
| Qwen3.8-27B by Qwen |
Huihui- from huihui-ai, 16.55 GB |
qwen35 |
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 | 32.9 tok/s, steady decode, 1 user, on TPU v6e-1 | mashbox run huihui-ai/Huihui-Qwen3.8-27B-abliterated-GGUF:Q4_K_M |
| Qwen3.8-27B by Qwen |
Qwen3.8- from 0bserverx, 16.55 GB |
qwen35 |
Q4_K_M | Quick check Runs on one chip of a TPU v6e-4 with the engine’s TPU kernels: short check against llama.cpp (4 prompts x 32 tokens) |
one chip of a TPU v6e-4 | 37.51 tok/s, steady decode, 1 user, on one chip of a TPU v6e-4 | mashbox run 0bserverx/Qwen3.8-27B-Heretic-Abliterated-Uncensored-GGUF:Q4_K_M |
| Qwen3.8-27B by Qwen |
Qwen3.8- from ggml-org, 18.97 GB |
qwen35 |
Q4_K_M | Quick check Runs on one chip of a TPU v6e-4 with the engine’s TPU kernels: short check against llama.cpp (4 prompts x 32 tokens) |
one chip of a TPU v6e-4 | 41.1 tok/s, steady decode, 1 user, on one chip of a TPU v6e-4 | mashbox run ggml-org/Qwen3.8-27B-GGUF:Q4_K_M |
| Qwen3.8-27B by Qwen |
Qwen3.8- from lmstudio-community, 16.81 GB |
qwen35 |
Q4_K_M | Quick check Runs on one chip of a TPU v6e-4 with the engine’s TPU kernels: short check against llama.cpp (4 prompts x 32 tokens) |
one chip of a TPU v6e-4 | 37.46 tok/s, steady decode, 1 user, on one chip of a TPU v6e-4 | mashbox run lmstudio-community/Qwen3.8-27B-GGUF:Q4_K_M |
| Qwen3.8-27B by Qwen |
Qwen3.8- from unsloth, 16.46 GB |
qwen35 |
Q4_K_M | Quick check Runs on one chip of a TPU v6e-4 with the engine’s TPU kernels: short check against llama.cpp (4 prompts x 32 tokens) |
one chip of a TPU v6e-4 | 32.13 tok/s, steady decode, 1 user, on one chip of a TPU v6e-4 | mashbox run unsloth/Qwen3.8-27B-GGUF:Q4_K_M |
| Qwen3.8-27B by Qwen |
Qwen3.8- from JonathanColetti, 16.81 GB |
qwen35 |
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 | 37.65 tok/s, steady decode, 1 user, on TPU v6e-1 | mashbox run JonathanColetti/Qwen3.8-27B-Uncensored-GGUF:Q4_K_M |
| Qwen3.8-27B by Qwen |
Qwen3.8- from HauhauCS, 17.92 GB |
qwen35 |
Q4_K_P | Quick check Runs on one chip of a TPU v6e-4 with the engine’s TPU kernels: short check against llama.cpp (4 prompts x 32 tokens) |
one chip of a TPU v6e-4 | 35.63 tok/s, steady decode, 1 user, on one chip of a TPU v6e-4 | mashbox run HauhauCS/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF:Q4_K_P |
| Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU by DavidAU |
Qwen3.8- from DavidAU, 18.5 GB |
qwen35 |
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 | 37.58 tok/s, steady decode, 1 user, on TPU v6e-1 | mashbox run DavidAU/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NEO-CODER-MAX-MTP-GGUF:Q4_K_M |
EXAONE
Section titled “EXAONE”| Model | GGUF file | Architecture | Quant | Level | Runs on | Speed | Run |
|---|---|---|---|---|---|---|---|
| EXAONE-3.5-2.4B-Instruct by LG AI Research |
EXAONE- from LGAI-EXAONE, 1.64 GB |
exaone |
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 | 284.75 tok/s, steady decode, 1 user, on TPU v6e-1 | mashbox run LGAI-EXAONE/EXAONE-3.5-2.4B-Instruct-GGUF:Q4_K_M |
Falcon
Section titled “Falcon”| Model | GGUF file | Architecture | Quant | Level | Runs on | Speed | Run |
|---|---|---|---|---|---|---|---|
| Falcon3-1B-Instruct by Technology Innovation Institute |
Falcon3- from bartowski, 1.78 GB |
llama |
Q8_0 | Quick check Runs on one chip of a TPU v6e-4 with the engine’s TPU kernels: short check against llama.cpp (4 prompts x 32 tokens) |
one chip of a TPU v6e-4 | 582.95 tok/s, steady decode, 1 user, on one chip of a TPU v6e-4 | mashbox run bartowski/Falcon3-1B-Instruct-GGUF:Q8_0 |
| Model | GGUF file | Architecture | Quant | Level | Runs on | Speed | Run |
|---|---|---|---|---|---|---|---|
| gemma-3-1b-it by Google DeepMind |
gemma- from unsloth, 0.81 GB |
gemma3 |
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 | 431.83 tok/s, steady decode, 1 user, on TPU v6e-1 | mashbox run unsloth/gemma-3-1b-it-GGUF:Q4_K_M |
Granite
Section titled “Granite”| Model | GGUF file | Architecture | Quant | Level | Runs on | Speed | Run |
|---|---|---|---|---|---|---|---|
| granite-3.3-2b-instruct by IBM |
granite- from ibm-granite, 1.55 GB |
granite |
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 | 231.85 tok/s, steady decode, 1 user, on TPU v6e-1 | mashbox run ibm-granite/granite-3.3-2b-instruct-GGUF:Q4_K_M |
| Model | GGUF file | Architecture | Quant | Level | Runs on | Speed | Run |
|---|---|---|---|---|---|---|---|
| Llama-3.2-1B-Instruct by Meta |
Llama- from bartowski, 0.81 GB |
llama |
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 | 465.05 tok/s, steady decode, 1 user, on TPU v6e-1 | mashbox run bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M |
| Llama-3.2-3B-Instruct by Meta |
Llama- from bartowski, 1.83 GB |
llama |
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 | 255 tok/s, steady decode, 1 user, on TPU v6e-1 | mashbox run bartowski/Llama-3.2-3B-Instruct-GGUF:IQ4_XS |
GGUF architectures with a passing chip test (general.architecture): qwen35 (Qwen3.5, Qwen3.6, Qwen3.8 (dense)), llama (Llama 3.x), qwen2 (Qwen2.5), qwen3 (Qwen3), gemma3 (Gemma 3), exaone (EXAONE 3.5), granite (Granite 3.3).
Listed, not in mashbox run
Section titled “Listed, not in mashbox run”Models we publish that mashbox run does not start in this version. Each names the device its files are made for and links the page that shows how to use them.
| Model | What it does | Made for | Files | How to use |
|---|---|---|---|---|
| Audio2Face-3D v3.0 by NVIDIA |
Speech audio to face motion for skin, tongue, jaw and eyes. | TPU | Audio2Face-3D-v3.0-TPU | How to use |
| Granite Speech 4.1 2B by IBM |
Speech recognition and speech translation. | TPU | granite-speech-4.1-2b-TPU | How to use |
| LFM3-1.7B by Liquid AI |
Text generation, as a Core ML model. | Apple Neural Engine | LFM3-1.7B-CoreML | How to use |
| maya1 and the SNAC 24 kHz decoder by Maya Research |
Text to speech from a voice description and the text to say. | TPU | maya1-TPU | How to use |
| Real-ESRGAN x2plus and x4plus by Xintao Wang |
Image upscaling. | TPU | Real-ESRGAN-TPU | How to use |
| Z-Image-Turbo by Tongyi-MAI |
Text to image. | TPU | Z-Image-Turbo-TPU | How to use |
Levels
Section titled “Levels”| Level | What the row means |
|---|---|
| MASHBOX build | One of our own TPU builds, with weights on Hugging Face and its own model page. |
| Tuned | A file that passed its chip test on the chips in its row, with a part tuned for it. |
| Tested | A file that passed its chip test on the chips in its row, with this version of mashbox. |
| Quick check | A file that ran on the chip in its row with this version of mashbox and matched llama.cpp in a short check: 4 prompts, 32 greedy tokens each. The row prints that check and the decode speed it measured for one request. |
| Checked on CPU | A file that passed its check against llama.cpp on CPU, with no passing chip test in this version of mashbox. The row prints the numbers from that CPU check. |
A GGUF file with no row here can still run: mashbox inspect FILE says whether it runs here, and why not, before anything loads. GGUF files says what runs.
Hugging Face safetensors
Section titled “Hugging Face safetensors”Some Hugging Face repos with safetensors weights, or a folder that holds one, run too. mashbox inspect OWNER/NAME reads the repo’s config and file headers and says whether and how the model would run on each chip, before any weights download.
mashbox run Qwen/Qwen3-4B --prompt "What is the capital of Australia, and why is it not Sydney?"This path reads dense text models of the llama, Mistral, Qwen2, Qwen3 and Qwen3.5 architectures, on one chip. Other architectures, routed experts, quantized weights and a model that fits no chip stop with a one-line reason. The command above ran on TPU v6e-1 with this version, with Qwen3-4B’s bfloat16 weights as published; it is the one model this path has run on a TPU chip with this version. The other architectures are checked on CPU against transformers on small test models.