Full specs

Hardware, the models it ships with, how fast they run, and how it stays up to date. Each figure links to where it comes from. Hardware describes the Boxes we ship today and can change between batches.

Hardware

One chip does the work. Memory is the only choice you make, and it can’t be changed after the order.

Processor: AMD Ryzen AI Max+ 395¹16 Zen 5 cores, 32 threads, 3.0 GHz base, up to 5.1 GHz
Graphics: Radeon 8060S¹40 compute units, up to 2.9 GHz
Memory: 64 GB or 128 GB¹LPDDR5x-8000 on a 256-bit bus, shared by the processor and graphics. Soldered, so it can’t be upgraded later.
Memory bandwidth: 256 GB/s¹Theoretical; we measure about 230 GB/s in use
Storage: 1 TB NVMe SSD¹On both memory options
Networking: Wi-Fi 7 and 5 Gbit Ethernet¹
Ports (rear): HDMI, 2 × DisplayPort, 2 × USB-C (USB4), 2 × USB-A, Ethernet, 3.5 mm audio¹
Power: 400 W internal supply, 80+ Gold¹Processor runs at 120 W sustained, 140 W boost
Size and weight: 4.5 litres, 3.1 kg¹96.8 × 205.5 × 226.1 mm

Models

Installed and tuned before the Box ships. More arrive through the Model Store.

Qwen 3.6 35B-A3B: Installed, on 64 GB and 128 GB²35B parameters, 3B active per token. The default. Quick enough for back-and-forth, and the one the speed figures below were measured on.
Gemma 4 26B-A4B: Installed, on 64 GB and 128 GB²26B parameters, 4B active per token. A second everyday model, and for short and mid-length prompts the quickest on the Box at reading what you give it.
Qwen 3.6 27B: Installed, on 64 GB and 128 GB²27B parameters, all active. Slower and more thorough. The one to pick for your hardest questions, when the answer matters more than the wait.
Qwen 3.8 27B: Model Store download, on 64 GB and 128 GB²27B parameters, all active. A newer model in the same vein, approved for the Box in August 2026.
Qwen 3.8 Flash-Next: Model Store download, on 128 GB only²125B parameters, 6B active per token. The largest model on the Box. Needs 128 GB of memory and the 0.2.8 update.

Speed

Measured on the 64 GB Box with the default model, Qwen 3.6 35B-A3B, in July 2026. A token is a word-piece.

Writing, everyday chat: 70+ tokens/s³Writing replies in everyday chat
Writing, 128K tokens in: 45+ tokens/s³Still writing 128K tokens into a long conversation
Reading documents: ~900 tokens/s³Reading the documents and notes you hand it
On the 128 GB Box: A single chat runs as fast as on 64 GB⁴Same chip, same memory bandwidth. One test run agrees; we haven’t run enough to quote a number, and speed with several chats at once is unmeasured on either size.

People at once

How many chats or agents run side by side, each with its own long context.

64 GB: 2 people or agents at once⁵Each with up to 256K tokens of context
128 GB: 4 people or agents at once⁵Each with up to 256K tokens of context

Software and security

Operating system: Fedora Linux⁶
Updates: Signed and checked on the Box before they install⁶If an update fails its health check, the Box rolls back to the version that worked.
Reaching it: A browser on your network, Tailscale from anywhere, or Signal^7,6Nothing is exposed to the internet until you turn remote access on.
Backups: Encrypted⁶

Price and support

Price: $3,300 with 64 GB, $4,800 with 128 GB⁹Both with a 1 TB SSD. US only.
Support: Two weeks of onboarding, then 12 months of technical support⁸By appointment, after setup and testing that happen before it ships.

See pricing

Sources and dates · checked 2 October 2026

1.Processor and graphics figures from AMD’s Ryzen AI Max+ 395 product page; chassis, I/O, power and size from the platform maker’s published spec sheet, both read 2 October 2026. The effective memory bandwidth of about 230 GB/s is our own measurement. Source
2.Factory model list for every shipped Box and the live Model Store catalogs, checked 5 October 2026. Every model runs on both memory sizes except Qwen 3.8 Flash-Next, a Model Store download that needs 128 GB and the 0.2.8 update.
3.Arx benchmark on the 64 GB Box, 15 July 2026: Qwen 3.6 35B-A3B (the default model), Q5 quantization, multi-token prediction on, prose prompts, served the way owners use it. Everyday speed was 72–75 tokens/s at the start of a chat, 45.6–48.2 tokens/s 128K tokens deep, and documents were read at 925–931 tokens/s with 8K tokens of context (a fresh 18K-token chat read at 786).
4.Arx reading on a 128 GB machine with the Box’s chip, 15 September 2026: the default model, in a single chat, ran as fast as on the 64 GB Box in the same real-chat test, as expected from the same chip and memory bandwidth. It was a single run, so we quote no 128 GB number, and speed with several chats at once has not been measured.
5.Box configuration as shipped: Qwen 3.6 35B-A3B and Gemma 4 run 2 × 256K-token slots on the 64 GB Box and 4 × 256K on the 128 GB Box. 256K is the models’ native context (262,144 tokens). Speed per person with several people at once has not been measured.
6.Arx update system: every update is Ed25519-signed and checked on the Box before it installs, with automatic rollback if the post-install health check fails (proven through the real update path on 25 September 2026). The Box runs Fedora Linux. Nothing on it is exposed to the internet by default: remote access is off until you turn it on. Backups are encrypted.
7.How owners reach the Box: in a browser on the same network, from a phone or another computer anywhere over a private Tailscale link, or by Signal message. Connecting these is part of onboarding support.
8.Terms of Sale §4: we handle setup, configuration, optimization and testing before the Box ships. After delivery, the first two weeks are onboarding support (weekdays 9–5 ET, by appointment, including AI configuration and automations); the first 12 months include free technical support by appointment. Source
9.Arx price list, 2 October 2026: $3,300 with 64 GB and $4,800 with 128 GB, both with a 1 TB SSD, US only, set up and supported.