Dual custom Ampere compute
Two modified Nvidia Ampere GPUs tuned as one compact local inference appliance.
Available for pre-order
Duet AI
A compact local AI appliance for builders who want dual custom-modded Ampere compute, private data, and high-VRAM workflows without assembling a workstation.
40GB
total GPU memory
1.5 TB/s
aggregate memory bandwidth
<400W load
~20W idle
USB-C + Wi-Fi
local terminal access
Finished hardware, not a hobby rig
A quiet desktop shell around a tuned dual custom Ampere core.
Duet AI is built for local model work that repeats all day: coding agents, document workflows, longer context experiments, and private inference. The enclosure keeps its custom-modded Ampere hardware controlled, cooled, and accessible through clean rear power, USB-C, and a direct development environment.
- CUDA local inference
- Native Nvidia acceleration for common local AI stacks.
- Thermal-first enclosure
- Large side and top ventilation for sustained sessions.
- Single pre-order configuration
- One tuned hardware package, one public pre-order price.
Custom Ampere architecture
Built on dual custom-modded Ampere GPUs.
40GB total VRAM
More working memory for larger quantized models, longer contexts, and repeatable local runs.
Designed for sustained inference
Custom cooling and power tuning keep the appliance desk-friendly under active local workloads.
Average score by model
Per-model averages against RTX 3090.
Duet AI average versus RTX 3090 average.
Across the listed model runs.
Gemma and Qwen benchmark averages.
| Model | Duet AI average | RTX 3090 average | Difference |
|---|---|---|---|
| Gemma 4 26B Q4/Q5 | 79 t/s | 88 t/s | 9 t/s lower |
| Qwen 3.6 27B Q4/Q5 | 72 t/s | 80 t/s | 8 t/s lower |
| Qwen 3.6 35B Q4/Q5 | 58 t/s | 64 t/s | 6 t/s lower |
Pre-order specification
One configuration, tuned for local inference.
GPU memory
40GB total VRAM
Bandwidth
Up to 1.5 TB/s aggregate
Access
USB-C, Wi-Fi, terminal workflow
Power
<400W full load, ~20W idle
Cooling
Multi-fan airflow path with top and side vents
Storage
1TB internal storage
Connection
USB-C to PC. Wi-Fi on network. Linux terminal access.
Connect by USB-C
Plug Duet AI into a PC and open the local terminal workflow.
Add it to Wi-Fi
After setup, connect it to the local network for cable-free access.
Run Linux locally
Duet AI runs an x86 Linux-based environment for local inference stacks.
Built and shipped
Handled by the Duet AI team.
Assembly, setup validation, packaging, and production-slot allocation are handled directly by the team behind the product.
Benchmark videos
YouTube benchmark videos.
Qwen 3.6 35B A3B
UD-Q6_K_XL local inference run
https://youtu.be/WHRsyEE-FRQ
Qwen 3.6 27B
Q8_0 local inference run
https://youtu.be/rTVFknP98Cg
Technical questions
System access and compatibility.
What is the internal control machine?
Duet AI includes an internal Intel N100 x86 system running Linux. You can update the Linux image or replace the operating system.
Are the GPU drivers compatible?
Yes. The GPU interface is not structurally changed, so Duet AI uses the native Nvidia driver stack and remains 100% compatible with standard drivers.
How does USB-C network access work?
Duet AI exposes only USB-C. Internally, the 2.5GbE Ethernet link is converted to USB-C, so it still behaves like a standard local TCP/IP network link with DHCP.
Is Duet AI compatible with macOS, Windows, and Linux?
Yes. Duet AI can be accessed from macOS, Windows, and Linux through the local USB-C network connection.
Why use an Ampere architecture?
Ampere delivers strong local inference performance at a lower hardware cost. The goal is to make quality local inference more accessible.
Why modify the GPUs?
The GPU modifications increase available VRAM and improve local inference performance for larger quantized models.
How is quality control handled?
Each unit goes through a substantial test battery over several consecutive days before shipment.
Is there a warranty?
Yes. Duet AI products are covered by a 1-year warranty starting from the order date.
Indiegogo campaign
Indiegogo campaign starts June 25, 2026.
The campaign opens on June 25, 2026. Direct pre-orders remain available now.
Public Indiegogo campaign opening date.
Orders can already be placed through the direct checkout.
40GB local inference appliance with 1TB internal storage.
Super Early Bird pre-order
Pre-order Duet AI at $1,585.
Deliveries have started. Super Early Bird pre-orders are allocated by production slot.
Offer
Super Early Bird
Price
$1,585 $1,785
Included
Duet AI, power cable, and user manual