Hardwarelattepandasingle-board-computerslocal-aiintelopenvinoedge-ai
DFRobot's $2,599 LattePanda Sigma 2 Puts an Intel Core Ultra X7, 64GB of RAM and a Preloaded Qwen Model on a 146mm Board
DFRobot opened orders for the LattePanda Sigma 2, a 3.5-inch board with Intel's Core Ultra X7 358H, 64GB of LPDDR5X and two 10GbE ports that ships with Ubuntu, OpenVINO Model Server and a 35-billion-parameter Qwen model ready to serve on the local network.
DFRobot sells it as a server, and the port list backs that up. There are two 10GbE jacks, two Thunderbolt 4 ports, an OCuLink port carrying PCIe 4.0 ×2 for an external GPU or storage, three M.2 M-key slots plus an E-key slot, and an RS-232/485 header for industrial gear. A Raspberry Pi RP2350 microcontroller sits beside the Intel chip and runs MicroPython, for the sensor and GPIO work an x86 CPU is clumsy at.
The memory is the headline for AI work. The 64GB is LPDDR5X-8533 shared between CPU and GPU, and DFRobot says “Up to 59GB of this unified memory can be allocated for GPU use.” That’s what lets a mixture-of-experts model like Qwen’s 35B-A3B fit at INT4 without a discrete graphics card.
The LattePanda quick start lists exactly what’s preinstalled: OpenVINO Model Server serving the model “through an OpenAI-compatible API (port 8080),” a base URL of http://localhost:8080/v3 with no API key required, and OpenClaw, “a pre-configured local AI agent with a browser Dashboard (port 18789).” The model files live in ~/models/. CNX Software, which went through the documentation, puts the platform at 180 TOPS combined and reports the preloaded model running at roughly 30 tokens per second, and says the $2,599 price runs until mid-November.
| LattePanda Sigma 2 | Previous LattePanda Sigma (per DFRobot) | |
|---|---|---|
| Processor | Intel Core Ultra X7 358H, 16 cores, up to 4.8GHz | not compared |
| Graphics | Arc B390, 12 Xe-cores | not compared |
| Memory | 64GB LPDDR5X-8533, up to 59GB to GPU | not compared |
| Idle power (average DC input) | 4.5W | 5W |
| Load power | 38W | 35W |
| Price | $2,599 pre-order, then $2,799 | not listed |
DFRobot’s own comparison stops at power. CNX names rivals built on the same class of Intel chip, the UP Xtreme PTL Edge, ADLINK’s COM-HPC-mPTL module and ASRock Industrial’s NUC(S) Ultra 300, though none of those ship with a model ready to serve.
What DFRobot doesn’t publish is just as useful to know. The product page gives no token-per-second figure of its own, no thermal numbers for long runs, and no word on how long Ubuntu and the preloaded model will get updates. The API needs no key out of the box, so anyone who can reach port 8080 on that network can use it.