A compact desktop workstation built around AMD’s Ryzen AI Max+ 395: sixteen Zen 5 cores, Radeon 8060S integrated graphics and an NPU, with 128 GB of LPDDR5X memory soldered to the board and shared by all three. There is no separate graphics card and no separate VRAM.

That single pool is the reason this class of machine exists. A discrete card is capped by its own 16 or 24 GB, so a large model has to be quantised hard, split across machines or rented elsewhere. Here the model sits in the same memory the processor addresses, and nothing has to be copied across a PCIe bus.

What it can do

• Run local inference on models that do not fit comfortably in a consumer graphics card.
• Serve a chat or coding assistant on your own network, with your own documents indexed for retrieval.
• Fine-tune smaller models and adapters without booking time on someone else’s GPU.
• Generate images in batches, and run several agents at once with separate context each.
• Keep working with the internet down: the weights are on the machine, not behind an API.
• Stay quiet and small enough for a desk rather than a rack.

What comes with it

• The machine itself, assembled in the configuration described on this page.
• The manufacturer’s power supply and quick-start documentation.
• An invoice from PureDNA, sent by e-mail.

What does not

• Monitor, keyboard, mouse or speakers. It is a desktop machine and needs a screen and mains power.
• An operating system licence. No OS is installed or licensed by PureDNA; the processor vendor validates the platform for Windows 11 64 bit and for current Linux distributions, and you install what you use.
• Extra memory. It is soldered to the board and cannot be expanded or replaced later.
• Storage beyond the capacity confirmed for your unit. The drive size is fixed when an order is placed.
• CUDA. There is no NVIDIA hardware here, so software that only runs on CUDA has to go through ROCm, Vulkan or the processor instead.

Before you order

Nothing on this page is a promise about delivery time, warranty or returns — those terms are written on How it works. This model is shown as out of stock and cannot be put in the cart.

Specifications

221 mm wide and 56 mm high — a little wider than an A4 sheet of paper (210 mm) and about the height of two stacked paperbacks.

Ports, numbered

The picture below is the rear panel of the YO1. The numbers match the list underneath, from left to right.

1. USB 2.0 (Type-A)
2. USB 2.0 (Type-A)
3. HDMI 2.1, 8K at 60 Hz
4. DisplayPort 1.4, 8K at 60 Hz
5. USB4 40 Gbps (Type-C)
6. USB 3.2 Gen2 (Type-A)
7. 2.5 Gigabit Ethernet (RJ45)
8. 3.5 mm headphone/microphone combo jack
9. DC power input

The front panel carries the power button, the performance-mode button, two USB 3.2 Gen2 Type-A ports, one USB4 Type-C port, the SD card reader and the second audio jack.

A compact desktop workstation built around AMD’s Ryzen AI Max+ 395 with 128 GB of LPDDR5X unified memory. Identical compute and memory envelope to the YO1; what changes is the enclosure, and with it the port layout, the expansion options and how the machine sheds heat.

Two configurations with the same silicon are not the same product. The YO2 is the one to look at if you are stacking several units, sitting one on a shelf, or need a particular set of ports at the back.

What it can do

• Run local inference on large models, with the model resident in unified memory rather than split across cards.
• Serve a coding or writing assistant on your own network and keep the data on your side of the connection.
• Fine-tune small models and adapters locally, and generate images in batches.
• Run several agents side by side, each with its own context.
• Work offline, because the weights live on the machine.

No acoustics or sustained-clock figures are published here. Cooling depends on placement and load, and is measured on the unit in front of you rather than read off a datasheet.

What comes with it

• The machine itself, assembled in the configuration described on this page.
• The manufacturer’s power supply and quick-start documentation.
• An invoice from PureDNA, sent by e-mail.

What does not

• Monitor, keyboard, mouse or speakers. It is a desktop machine and needs a screen and mains power.
• An operating system licence. No OS is installed or licensed by PureDNA; the processor vendor validates the platform for Windows 11 64 bit and for current Linux distributions, and you install what you use.
• Extra memory. It is soldered to the board and cannot be expanded or replaced later.
• Storage beyond the capacity confirmed for your unit. The drive size is fixed when an order is placed.
• CUDA. There is no NVIDIA hardware here, so software that only runs on CUDA has to go through ROCm, Vulkan or the processor instead.

Before you order

Nothing on this page is a promise about delivery time, warranty or returns — those terms are written on How it works. This model is shown as out of stock and cannot be put in the cart.

Specifications

Ports, numbered

The rear-panel photograph and the numbered connector list for the YO2 are still with the manufacturer; they go onto this page as soon as the drawing arrives. The photographs above show the front panel. What is confirmed today is listed in the table, and nothing beyond it is claimed here.

The TXA1 with 64 GB of LPDDR5X unified memory: same sixteen-core Ryzen AI Max+ 395 and Radeon 8060S graphics as the 128 GB version, with half the shared pool. Everything the processor, the graphics unit and the NPU touch comes out of that one 64 GB.

What it can do

• Run local inference on mid-size models — roughly up to 30B parameters at 4-bit quantisation, in practice less once context is included.
• Serve a coding assistant or a document search over your own files on a local network.
• Generate images, and run a couple of agents at the same time.
• Work offline, with the weights on your own disk.

When the 128 GB version is the better buy

If you intend to run 70B-class models, keep several model sets loaded at once, or run more than two agents in parallel, take a 128 GB configuration instead. The memory is soldered: there is no upgrade path afterwards, and the difference cannot be bought later as a part.

What comes with it

• The machine itself, assembled in the configuration described on this page.
• The manufacturer’s power supply and quick-start documentation.
• An invoice from PureDNA, sent by e-mail.

What does not

• Monitor, keyboard, mouse or speakers. It is a desktop machine and needs a screen and mains power.
• An operating system licence. No OS is installed or licensed by PureDNA; the processor vendor validates the platform for Windows 11 64 bit and for current Linux distributions, and you install what you use.
• Extra memory. It is soldered to the board and cannot be expanded or replaced later.
• Storage beyond the capacity confirmed for your unit. The drive size is fixed when an order is placed.
• CUDA. There is no NVIDIA hardware here, so software that only runs on CUDA has to go through ROCm, Vulkan or the processor instead.

Before you order

Nothing on this page is a promise about delivery time, warranty or returns — those terms are written on How it works. This model is shown as out of stock and cannot be put in the cart.

Specifications

Ports, numbered

The picture below is the rear panel. The numbers match the list underneath, from left to right.

1. 2 × Wi-Fi antenna connector (SMA)
2. 3.5 mm audio jack
3. 2.5 Gigabit Ethernet (RJ45)
4. 10 Gigabit Ethernet (RJ45)
5. 4 × USB 3.2 Gen1 (Type-A, 5 Gbps)
6. DisplayPort 2.1
7. HDMI 2.1
8. USB4 40 Gbps (Type-C)
9. DC power input

The front panel carries the power button, two USB 3.2 Gen2 Type-A ports, one USB4 Type-C port and the SD card reader.

A small workstation on AMD’s Ryzen AI Max+ 395 with 128 GB of LPDDR5X unified memory, four M.2 NVMe slots, 2.5 GbE and 10 GbE Ethernet, and Wi-Fi 7.

The combination is the point. 128 GB of shared memory decides which models you can load; four M.2 slots decide how many of them you can keep on disk at once; 10 GbE decides how quickly datasets and checkpoints move to the next machine. One of those alone is a specification, all three together are a working local stack.

What it can do

• Run large models locally, resident in one memory pool instead of split across cards.
• Keep several model sets on separate drives and switch between them without re-downloading.
• Serve inference to other machines on a 10 GbE network, and move checkpoints between them.
• Fine-tune small models and adapters, and generate images or short video locally.
• Run Windows 11 or Linux; dual boot is supported at platform level. Driver and ROCm support change with every release, so check the current state for the framework you intend to use.

About the four M.2 slots

The slots are there; the drives are not, beyond the capacity confirmed for your unit. That is a decision for you rather than a fixed part of the price, and it means storage can be expanded later — unlike the memory.

What comes with it

• The machine itself, assembled in the configuration described on this page.
• The manufacturer’s power supply and quick-start documentation.
• An invoice from PureDNA, sent by e-mail.

What does not

• Monitor, keyboard, mouse or speakers. It is a desktop machine and needs a screen and mains power.
• An operating system licence. No OS is installed or licensed by PureDNA; the processor vendor validates the platform for Windows 11 64 bit and for current Linux distributions, and you install what you use.
• Extra memory. It is soldered to the board and cannot be expanded or replaced later.
• Storage beyond the capacity confirmed for your unit. The drive size is fixed when an order is placed.
• CUDA. There is no NVIDIA hardware here, so software that only runs on CUDA has to go through ROCm, Vulkan or the processor instead.

Before you order

Nothing on this page is a promise about delivery time, warranty or returns — those terms are written on How it works. This model is shown as out of stock and cannot be put in the cart.

Specifications

Ports, numbered

The picture below is the rear panel. The numbers match the list underneath, from left to right.

1. 2 × Wi-Fi antenna connector (SMA)
2. 3.5 mm audio jack
3. 2.5 Gigabit Ethernet (RJ45)
4. 10 Gigabit Ethernet (RJ45)
5. 4 × USB 3.2 Gen1 (Type-A, 5 Gbps)
6. DisplayPort 2.1
7. HDMI 2.1
8. USB4 40 Gbps (Type-C)
9. DC power input

The front panel carries the power button, two USB 3.2 Gen2 Type-A ports, one USB4 Type-C port and the SD card reader.

The largest configuration in this category. Ryzen AI Max+ PRO 495 with sixteen cores and 32 threads, Radeon 8065S integrated graphics with 40 RDNA 3.5 compute units, and 192 GB of LPDDR5X-8533 unified memory. Up to 160 GB of that memory can be handed to the graphics unit.

The memory is the product. It is soldered, fast and shared: the 160 GB the graphics unit holds is memory the processor stops using while it holds it. One pool, not two — that is what makes models in the 300B class at 4-bit quantisation practical on a desk, and it is also the constraint you plan around.

What it can do

• Load models that need more than 128 GB, including large mixture-of-experts checkpoints at 4-bit.
• Run several agents side by side, each with a separate KV cache and long context.
• Generate video and image batches locally instead of renting a GPU by the hour.
• Fine-tune mid-size models on data that never leaves the machine.
• Serve inference over the network to a small team.

Validated at processor level for Windows 11 64-bit, Ubuntu and RHEL.

What comes with it

• The machine itself, assembled in the configuration described on this page.
• The manufacturer’s power supply and quick-start documentation.
• An invoice from PureDNA, sent by e-mail.

What does not

• Monitor, keyboard, mouse or speakers. It is a desktop machine and needs a screen and mains power.
• An operating system licence. No OS is installed or licensed by PureDNA; the processor vendor validates the platform for Windows 11 64 bit and for current Linux distributions, and you install what you use.
• Extra memory. It is soldered to the board and cannot be expanded or replaced later.
• Storage beyond the capacity confirmed for your unit. The drive size is fixed when an order is placed.
• CUDA. There is no NVIDIA hardware here, so software that only runs on CUDA has to go through ROCm, Vulkan or the processor instead.

Before you order

Nothing on this page is a promise about delivery time, warranty or returns — those terms are written on How it works. This model is shown as out of stock and cannot be put in the cart.

Specifications

180 mm wide and 78 mm high — a little narrower than an A4 sheet of paper (210 mm) and about the height of a paperback standing on its edge.

Ports, numbered

The photographs above show the rear panel from the side. The flat rear-panel drawing with numbered connectors is still with the manufacturer and goes onto this page when it arrives; the port set is listed in the table under Wired network, Video outputs and USB.