Decision ModelsDecision Models

Pre-installed hardware · Coming soon

Run decision models on your own hardware.

From a small edge board to a rack of GPUs: see which models fit, how many decisions per second to expect, and what each machine costs. We are preparing a pilot program for pre-installed machines. Register interest in a device below; we will confirm feasibility and availability before offering a quote. You can already use your own device with our installer.

Compare devicesBecome a pilot customer

Why self-host decisions

Decisions are a good fit for your own hardware

01

Sensitive

Decisions are made on customer records, contracts, transactions and internal policies. On your own machine, that data never leaves your network.

02

Latency-critical

A decision often sits inside a user request or a control loop. A local model answers without a round trip to a hosted API.

03

High-volume

Classification, routing and review run on every item, all day. A machine you own has a fixed cost however many decisions it makes.

04

Offline capable

After the model files are downloaded, the machine can serve decisions without any internet connection.

Why it runs fast

The REAX Engine

Decision models read a prompt and score a few answer options. They do not write long text. The REAX Engine is our serving engine built for exactly that: it reads the prompt once, scores the options and skips token-by-token generation.

On one NVIDIA RTX PRO 6000, Perplexity Decider v1.1 went from 5.0 decisions per second on the author’s reference server to 15.7 on the REAX Engine (FP8), with the same benchmark scores within noise.

The REAX Engine runs on NVIDIA GPUs (CUDA). On Apple Silicon and other platforms we plan to use standard open-source runtimes (not tested yet).

Models in this guide

What we size for

  • s1-pro class · 26–27B dense: Perplexity Decider v1.1 (27B), vLLM-SR d3 (26B)
  • 26B-A4B MoE · Blink v0.3 (s1-fast), Surogate Rune
  • Quyet-1.0-Large · 31B dense
  • 4B models · e.g. H2O-Lightning-4B, Plumb-4B

Memory figures assume the precision shown in the table. Lower precision uses less memory but must pass our quality checks before we use it. Blink v0.3 for local installs is offered subject to its licence.

Selected model: Fastino GLiDE 1.2

Check the memory it needs on the model page, then pick a device below with at least that much memory.

View model requirements →

Device tiers

Edge, desk, workstation, rack

Prices are device street or list prices in Germany (incl. VAT) and the US, with the date we checked them. Setup, shipping and support are quoted separately on request.

Tier 1

Edge

Small, low-power boards next to a machine, a camera or a robot. Good for 4B models; the 128 GB Thor also holds the larger MoE models.

Illustration of a GPU edge board; not a product photoComing soon · Pilot interest

NVIDIA Jetson Orin Nano Super Developer Kit

Memory
8 GB
Price
€559 (Voelkner) · $399
Power
7–25 W
Noise
Small fan; no dB figure published
  • s1-pro class: No Does not fit
  • 26B-A4B MoE: No Does not fit
  • Quyet-1.0-Large: No Does not fit
  • 4B models: Fits 4-bit / 8-bit fits · speed not estimated

Bring your own board; our installer and setup guide. Not yet tested by us.

Sources (3)
Interested in becoming a pilot customer
Illustration of a GPU edge board; not a product photoComing soon · Pilot interest

NVIDIA Jetson AGX Thor Developer Kit

The Jetson AGX Orin 64 GB is the smaller option for 4B and 8-bit MoE models.

Memory
128 GB unified
Price
€8,010 (Generation Robots) · $5,499
Power
40–130 W module
Noise
Active cooling; no dB figure published
  • s1-pro class: No Not yet: REAX Engine builds are x86 NVIDIA only today
  • 26B-A4B MoE: Estimate BF16 · ≈7–13 dec/s (estimate)
  • Quyet-1.0-Large: Fits BF16 fits in memory · speed not estimated
  • 4B models: Estimate BF16 · ≈15–25 dec/s (estimate)

Bring your own kit; our installer and setup guide. Not yet tested by us.

Sources (4)
Interested in becoming a pilot customer

Tier 2

Desk

Quiet boxes that sit on a desk or in a small office rack. The 128 GB GB10 systems hold every model in this lineup at full precision.

Illustration of a mini PC; not a product photoComing soon · Pilot interest

NVIDIA GB10 systems (DGX Spark class)

ASUS Ascent GX10, Gigabyte AI TOP ATOM, Dell Pro Max with GB10, Lenovo ThinkStation PGX, HP ZGX Nano, NVIDIA DGX Spark

Memory
128 GB unified
Price
€6,468–6,491 in stock (Alternate, 12 Oct) · DGX Spark $6,950
Power
240 W power supply
Noise
Quiet desktop box (NVIDIA lists about 29 dB under load)
  • s1-pro class: Estimate FP8 on REAX Engine · ≈2–4 dec/s (estimate; needs an Arm build, not yet tested)
  • 26B-A4B MoE: Estimate BF16 / NVFP4 · ≈15–25 dec/s (estimate)
  • Quyet-1.0-Large: Estimate NVFP4 · ≈5–10 dec/s (estimate)
  • 4B models: Estimate BF16 · ≈30–50 dec/s (estimate)

Pre-install planned on the OEM-branded boxes (ASUS, Gigabyte, Dell, Lenovo, HP) once our Arm (aarch64) builds pass testing. Until then: bring your own.

Sources (5)
Interested in becoming a pilot customer
Illustration of a mini PC; not a product photoComing soon · Pilot interest

Apple Mac mini with M5 Pro, 64 GB

Memory
64 GB unified
Price
from €3,099 · $2,699 (Apple list price at launch)
Power
6 W idle · 145 W max
Noise
No dB figure published
  • s1-pro class: No Not yet: needs the REAX Engine (NVIDIA CUDA)
  • 26B-A4B MoE: Fits 8-bit fits in memory · Apple Silicon build not tested yet
  • Quyet-1.0-Large: Fits 8-bit fits in memory · Apple Silicon build not tested yet
  • 4B models: Fits BF16 fits · Apple Silicon build not tested yet

Pre-install planned once our Apple Silicon builds pass testing (macOS stays untouched; our software is added as separate apps).

Sources (3)
Interested in becoming a pilot customer
Illustration of a mini PC; not a product photoComing soon · Pilot interest

AMD Ryzen AI Max+ 395 mini-PCs, 128 GB

Framework Desktop, HP Z2 Mini G1a, GMKtec EVO-X2

Memory
128 GB unified
Price
€3,889 (Framework DE configurator) – €6,084 incl. VAT (HP at Bechtle)
Power
About 120–145 W under AI load (reviews of similar boxes)
Noise
About 43–45 dBA under load (reviews of similar boxes)
  • s1-pro class: No Not yet: needs the REAX Engine (NVIDIA CUDA)
  • 26B-A4B MoE: Fits Fits in memory; GPU support in our installer is not ready yet
  • Quyet-1.0-Large: Fits Fits in memory; GPU support not ready yet
  • 4B models: Fits Fits; CPU recipes only today

Listed for comparison. We will add it once our installer supports AMD GPUs.

Sources (4)
Interested in becoming a pilot customer

Tier 3

Workstation

A tower under the desk or a large Mac. This is where we have our own measurements: the RTX PRO 6000 runs each model in the lineup on a single card.

Illustration of a workstation tower; not a product photoComing soon · Pilot interest

Apple Mac Studio with M5 Max or M5 Ultra

M5 Max up to 128 GB; M5 Ultra 96 or 256 GB (512 GB announced for late October)

Memory
96–256 GB unified (M5 Ultra)
Price
M5 Ultra from €6,599 · $5,499 (Apple list price at launch)
Power
M5 Max 7 W idle · 200 W max; M5 Ultra 9 W idle · 385 W max
Noise
Very quiet; fan audible only under sustained load (reviews)
  • s1-pro class: No Not yet: needs the REAX Engine (NVIDIA CUDA)
  • 26B-A4B MoE: Fits BF16 fits in memory · Apple Silicon build not tested yet
  • Quyet-1.0-Large: Fits BF16 fits (96 GB and up) · Apple Silicon build not tested yet
  • 4B models: Fits BF16 fits · Apple Silicon build not tested yet

Pre-install planned once our Apple Silicon builds pass testing (macOS stays untouched; our software is added as separate apps).

Sources (4)
Interested in becoming a pilot customer

Tier 4

Rack

Several GPUs in one chassis for high volume. Each GPU runs its own copy of a model, so throughput grows with the number of GPUs.

Illustration of a rack server; not a product photoComing soon · Pilot interest

tiny corp tinybox

New tinybox with 1–2 RTX PRO 6000 (from $7,000 without GPUs, +$19,000 per GPU); tinybox pro v2 with 8× RTX 5090 ($100,000) or 8× RTX PRO 6000 ($200,000)

Memory
96–768 GB GPU memory
Price
$7,000–200,000, sold in USD only; import VAT and duty extra
Power
1,700 W supply (tinybox); 4 × 2,000 W (pro v2)
Noise
Under 50 dB (tinybox); pro v2 is a 5U rack server
  • s1-pro class: Estimate About the single-GPU figure × number of GPUs (estimate)
  • 26B-A4B MoE: Estimate About the single-GPU figure × number of GPUs (estimate)
  • Quyet-1.0-Large: Estimate About the single-GPU figure × number of GPUs (estimate)
  • 4B models: Estimate About the single-GPU figure × number of GPUs (estimate)

Order directly from tiny corp; we install our software remotely over SSH.

Sources (2)
Interested in becoming a pilot customer

Model fit and speed

Which model fits where, and how fast

Decisions per second (dec/s) on Decision Index traffic, several questions per request, at the best tested concurrency. Short single-question prompts run faster; long documents run slower.

Devices1-pro class26B-A4B MoEQuyet-1.0-Large4B models
NVIDIA Jetson Orin Nano Super Developer Kit
Edge · 8 GB
No
Does not fit
No
Does not fit
No
Does not fit
Fits
4-bit / 8-bit fits · speed not estimated
NVIDIA Jetson AGX Thor Developer Kit
Edge · 128 GB unified
No
Not yet: REAX Engine builds are x86 NVIDIA only today
Estimate
BF16 · ≈7–13 dec/s (estimate)
Fits
BF16 fits in memory · speed not estimated
Estimate
BF16 · ≈15–25 dec/s (estimate)
NVIDIA GB10 systems (DGX Spark class)
Desk · 128 GB unified
Estimate
FP8 on REAX Engine · ≈2–4 dec/s (estimate; needs an Arm build, not yet tested)
Estimate
BF16 / NVFP4 · ≈15–25 dec/s (estimate)
Estimate
NVFP4 · ≈5–10 dec/s (estimate)
Estimate
BF16 · ≈30–50 dec/s (estimate)
Apple Mac mini with M5 Pro, 64 GB
Desk · 64 GB unified
No
Not yet: needs the REAX Engine (NVIDIA CUDA)
Fits
8-bit fits in memory · Apple Silicon build not tested yet
Fits
8-bit fits in memory · Apple Silicon build not tested yet
Fits
BF16 fits · Apple Silicon build not tested yet
AMD Ryzen AI Max+ 395 mini-PCs, 128 GB
Desk · 128 GB unified
No
Not yet: needs the REAX Engine (NVIDIA CUDA)
Fits
Fits in memory; GPU support in our installer is not ready yet
Fits
Fits in memory; GPU support not ready yet
Fits
Fits; CPU recipes only today
Workstation with NVIDIA RTX PRO 6000 Blackwell
Workstation · 96 GB GPU memory
Measured
FP8 on REAX Engine · 15.7 dec/s (measured)
Measured
Blink NVFP4 · 97 dec/s · Rune BF16 · 76 dec/s (measured)
Measured
FP8 · 28.7 dec/s (measured); NVFP4 · 36.7 (measured, still in quality review)
Measured
Plumb-4B BF16 · about 200 dec/s (measured, short single-question prompts)
Workstation with NVIDIA GeForce RTX 5090
Workstation · 32 GB GPU memory
Measured
FP8 on REAX Engine · 10.3 dec/s (measured, offline throughput)
Estimate
Blink NVFP4 · ≈60–80 dec/s (estimate; tight fit, shorter context); Rune BF16 does not fit
Estimate
NVFP4 · ≈25–30 dec/s (estimate)
Estimate
BF16 · ≈140–170 dec/s (estimate)
Apple Mac Studio with M5 Max or M5 Ultra
Workstation · 96–256 GB unified (M5 Ultra)
No
Not yet: needs the REAX Engine (NVIDIA CUDA)
Fits
BF16 fits in memory · Apple Silicon build not tested yet
Fits
BF16 fits (96 GB and up) · Apple Silicon build not tested yet
Fits
BF16 fits · Apple Silicon build not tested yet
tiny corp tinybox
Rack · 96–768 GB GPU memory
Estimate
About the single-GPU figure × number of GPUs (estimate)
Estimate
About the single-GPU figure × number of GPUs (estimate)
Estimate
About the single-GPU figure × number of GPUs (estimate)
Estimate
About the single-GPU figure × number of GPUs (estimate)

Swipe horizontally to see all models.

How we measure and estimate

Measured figures are our own runs on rented NVIDIA RTX PRO 6000 (96 GB) and RTX 5090 (32 GB) cards in October 2026. The RTX 5090 figure for s1-pro-class models is offline packed throughput from our hardware sweep.

Estimates scale our RTX PRO 6000 measurements by the ratio of published prompt-processing benchmarks between that card and the target device (LMSYS, JetsonHacks). They can be off by about half in either direction. We have not yet run our models on Apple Silicon, AMD, Jetson or GB10 hardware, so those devices show estimates or only a memory fit; we benchmark your configuration before delivery.

Fits means the model weights at that precision fit in the device memory with room for context; we do not give a speed where no comparable benchmark exists.

How we can deliver

Pre-installed, or bring your own

Coming soon · Pilot program

A ready machine

For pilot customers, we plan to source a new workstation through regular retail or distribution channels in the EU, install the runtime, model and a Jev-compatible endpoint, and run a burn-in test before shipping. Registering interest is not an order or a delivery commitment.

  • Pilot candidates: workstations with RTX PRO 6000 or RTX 5090
  • GB10 systems (ASUS, Gigabyte, Dell, Lenovo, HP), once our Arm builds pass testing
  • Mac mini and Mac Studio, once our Apple Silicon builds pass testing; macOS stays untouched and our software is added as separate apps
Bring your own device

Your machine, our installer

Use hardware you already own or buy directly from the maker. Our installer and setup guide detects the device, downloads pinned model files and starts a local endpoint. We can also install remotely over SSH.

  • NVIDIA Jetson boards and GB10 systems
  • tiny corp tinybox (sold directly by tiny corp)
  • Any Linux machine with a recent NVIDIA GPU

We are an independent integrator. We are not an authorised reseller or partner of Apple, NVIDIA, AMD, tiny corp or any system maker, and product names are used only to identify the hardware. The manufacturer’s warranty follows the manufacturer’s own terms; as the seller we provide the statutory warranty. Quotes are confirmed by email.

Cloud and remote GPUs

Use a GPU you already rent

For each model, this lists the least expensive sourced instance with enough recorded GPU memory. Prices change; check before renting. Prefer a managed server? See dedicated hosting.

ModelRecommended VRAMCheapest matching instancePriceSource and date
AutoJev-27B80 GBRunPod · A100 80GB PCIe secure
A100 · 80 GB
$1.79/hPrice source · 2026-10-09 · not stated (API)
basal-omni-1.0-ultra80 GBRunPod · A100 80GB PCIe secure
A100 · 80 GB
$1.79/hPrice source · 2026-10-09 · not stated (API)
Bobcat Flash 1.280 GBRunPod · A100 80GB PCIe secure
A100 · 80 GB
$1.79/hPrice source · 2026-10-09 · not stated (API)
Cygnet48 GBRunPod · L40S community
L40S · 48 GB
$0.79/hPrice source · 2026-10-09 · not stated (API)
decider-12b v264 GBRunPod · A100 80GB PCIe secure
A100 · 80 GB
$1.79/hPrice source · 2026-10-09 · not stated (API)
decisio v0.8.0 on gemma-4-31B-it48 GBRunPod · L40S community
L40S · 48 GB
$0.79/hPrice source · 2026-10-09 · not stated (API)
deck-31B80 GBRunPod · A100 80GB PCIe secure
A100 · 80 GB
$1.79/hPrice source · 2026-10-09 · not stated (API)
Fastino GLiDE 1.280 GBRunPod · A100 80GB PCIe secure
A100 · 80 GB
$1.79/hPrice source · 2026-10-09 · not stated (API)
H2O-Lightning-4B v1.124 GBRunPod · RTX 4090 community
RTX 4090 · 24 GB
$0.34/hPrice source · 2026-10-09 · not stated (API)
Imajev-4B12 GBRunPod · RTX 4090 community
RTX 4090 · 24 GB
$0.34/hPrice source · 2026-10-09 · not stated (API)
Jeff-1.0-Large80 GBRunPod · A100 80GB PCIe secure
A100 · 80 GB
$1.79/hPrice source · 2026-10-09 · not stated (API)
Jev-Omni48 GBRunPod · L40S community
L40S · 48 GB
$0.79/hPrice source · 2026-10-09 · not stated (API)
Jevify Gemma 4 26B-A4B64 GBRunPod · A100 80GB PCIe secure
A100 · 80 GB
$1.79/hPrice source · 2026-10-09 · not stated (API)
JPT-4B12 GBRunPod · RTX 4090 community
RTX 4090 · 24 GB
$0.34/hPrice source · 2026-10-09 · not stated (API)
KnowLine 31B-Gen180 GBRunPod · A100 80GB PCIe secure
A100 · 80 GB
$1.79/hPrice source · 2026-10-09 · not stated (API)
Mammoth-180 GBRunPod · A100 80GB PCIe secure
A100 · 80 GB
$1.79/hPrice source · 2026-10-09 · not stated (API)
NeoHorse Jev 4B12 GBRunPod · RTX 4090 community
RTX 4090 · 24 GB
$0.34/hPrice source · 2026-10-09 · not stated (API)
Perplexity Decider v1.1 27B80 GBRunPod · A100 80GB PCIe secure
A100 · 80 GB
$1.79/hPrice source · 2026-10-09 · not stated (API)
Quyet-1.0-Large80 GBRunPod · A100 80GB PCIe secure
A100 · 80 GB
$1.79/hPrice source · 2026-10-09 · not stated (API)
reflex 4B12 GBRunPod · RTX 4090 community
RTX 4090 · 24 GB
$0.34/hPrice source · 2026-10-09 · not stated (API)
René-1 31B FP848 GBRunPod · L40S community
L40S · 48 GB
$0.79/hPrice source · 2026-10-09 · not stated (API)
shisa-de-164 GBRunPod · A100 80GB PCIe secure
A100 · 80 GB
$1.79/hPrice source · 2026-10-09 · not stated (API)
Surogate Rune 26B-A4B v364 GBRunPod · A100 80GB PCIe secure
A100 · 80 GB
$1.79/hPrice source · 2026-10-09 · not stated (API)
torchcast-decision-12b48 GBRunPod · L40S community
L40S · 48 GB
$0.79/hPrice source · 2026-10-09 · not stated (API)
vjev-vision12 GBRunPod · RTX 4090 community
RTX 4090 · 24 GB
$0.34/hPrice source · 2026-10-09 · not stated (API)
vLLM-SR d380 GBRunPod · A100 80GB PCIe secure
A100 · 80 GB
$1.79/hPrice source · 2026-10-09 · not stated (API)
vLLM-SR d3-turbo80 GBRunPod · A100 80GB PCIe secure
A100 · 80 GB
$1.79/hPrice source · 2026-10-09 · not stated (API)
Winnow-12B Q816 GBRunPod · RTX 4090 community
RTX 4090 · 24 GB
$0.34/hPrice source · 2026-10-09 · not stated (API)
Xor 26B-A4B64 GBRunPod · A100 80GB PCIe secure
A100 · 80 GB
$1.79/hPrice source · 2026-10-09 · not stated (API)

Swipe horizontally to see all cloud price and source details.

Coming soon · Pilot interest

Interested in becoming a pilot customer?

Tell us the model, the expected volume and where the machine will run. We will check your chosen device and model, then reply about pilot availability. No payment or order is taken here; pricing and delivery would be agreed separately.

Models
Run decision models on your own hardware — Decision Models