LattePanda Sigma 2 Lets Its Intel Arc GPU Claim 59GB of 64GB Memory, on a $2,599 Board the Size of the Original Sigma

DFRobot calls the Sigma 2 “The Embeddable AI Single Board Server,” and the spec sheet backs up the server part. There are two 10Gb Ethernet jacks, an OCuLink port wired as PCIe 4.0 ×2, and three M-key M.2 slots, one of them PCIe 5.0 ×2. The store page lists the memory as “64GB dual-channel LPDDR5X-8533 unified memory, with up to 59GB allocatable for GPU use.”

That 59GB figure is the whole pitch. Local language models live or die on how much memory the GPU can reach, and a board that hands most of 64GB to its graphics side can hold a model a gaming card with 16GB can’t.

The software arrives already set up. LattePanda’s getting-started guide lists Ubuntu 26.04 with OpenVINO Model Server, which “serves the local AI model through an OpenAI-compatible API (port 8080),” plus OpenClaw, “a pre-configured local AI agent with a browser Dashboard (port 18789).” The model on the drive is Qwen3.6-35B-A3B in INT4. CNX Software’s Jean-Luc Aufranc reports that this model “runs locally at around 30 tokens per second,” and calls the price “not exactly cheap.”

There’s also an RP2350 microcontroller on the board with MicroPython preloaded, the same chip family as a Raspberry Pi Pico 2. It’s there for the relays-and-sensors side of a project, the job the ATmega32U4 did on the first Sigma.

Here’s how it lines up against the board it replaces and two other ways to buy a lot of GPU-reachable memory:

	LattePanda Sigma 2	LattePanda Sigma	Framework Desktop (Max+ 395, 64GB)	NVIDIA DGX Spark
Price	$2,599 pre-order, then $2,799	$709 to $928	$1,959	Not on NVIDIA’s page
Processor	Intel Core Ultra X7 358H, 16 cores	Intel i5-1340P, 12 cores	AMD Ryzen AI Max+ 395	Up to 1 petaFLOP FP4 with sparsity
Memory	64GB LPDDR5X-8533, up to 59GB for GPU	Up to 32GB LPDDR5-6400	64GB	128GB LPDDR5X
Networking	2 × 10GbE	2 × 2.5GbE	5Gbit Ethernet	10GbE plus ConnectX-7 at 200Gbps
Size	146 × 102mm board	146 × 102mm board	4.5L case	150 × 150 × 50.5mm
Comes with	1TB NVMe with Ubuntu and the AI stack	SSD only on some models	Not compared	DGX OS

What they didn’t say. DFRobot’s speed claim is relative: “nearly twice the text-generation speed of the previous-generation Sigma in the illustrated local AI test,” with no model or token count named on the store page. The 30 tokens per second figure comes from CNX, not DFRobot. There’s no ship date beyond pre-order, and no case in the box; DFRobot sells an aluminum one for $49 more. CNX also notes the company lists support for “Windows 11 or Ubuntu 24.04 LTS or later,” while the drive arrives with 26.04.

What it means for people who run systems. The getting-started guide prints the default login in plain text: username lattepanda, password lattepanda. Put that beside an OpenAI-style API on port 8080 and an agent dashboard on port 18789, and the first job after unboxing is a new password and a firewall rule, before the board goes anywhere near a shared network. The guide also warns that “the first request may take up to 30 seconds or more,” which matters if a health check is waiting on it.

Why now. Intel’s new Core Ultra X7 parts put a GPU with shared LPDDR5X on a mobile-class chip, so a board maker can now promise the same kind of large GPU memory that until recently meant an AMD Strix Halo desktop or NVIDIA’s DGX Spark.

The open question is the price gap. At $2,599 the Sigma 2 costs about $640 more than Framework’s 64GB desktop, and what that money buys is the small board, the 10GbE ports and a drive that’s already set up.