LattePanda announced the Sigma 2 in a PR Newswire release dated October 22, 2026, from Shanghai. The company calls it an "embeddable AI single board server." CNX Software published a detailed write-up on October 29. The board keeps the general idea of the original Sigma, an x86 machine on a board small enough to build into something, but the hardware is now aimed at running large language models locally. It uses Intel's Core Ultra X7 358H, takes up to 64GB of LPDDR5X, and has three M.2 M-Key slots, two 10GbE ports and an RP2350 microcontroller that handles the GPIO header.

The board is on pre-sale now. LattePanda's release prices it at $2,599 through mid-November, after which it returns to $2,799. None of the sources give a shipping date.

The SoC, and what "180 TOPS" actually counts

LattePanda's product page lists the Core Ultra X7 358H as a 16-core, 16-thread part with clocks up to 4.8 GHz. CNX Software gives the core split as 4 performance cores, 8 efficiency cores and 4 low-power efficiency cores (4P+8E+4LPE), with 18MB of cache and a Processor Base Power of 25 W. The integrated GPU is a 12 Xe-core Arc B390.

The 180 TOPS headline needs some explanation. LattePanda's product page describes it as "overall peak INT8": 122 TOPS from the GPU, 50 TOPS from the NPU and 8 TOPS from the CPU. Those are three separate engines added together. LattePanda itself says the compute is not combined for a single workload. A single model running on the iGPU should therefore be sized against the 122 TOPS GPU figure, not 180. For anyone sizing an inference job, the GPU and NPU numbers are the ones that matter, and all of these are peak INT8 ratings, not sustained throughput.

Memory is the real story

For local LLM work, memory bandwidth and capacity usually limit you before TOPS do. The product page lists up to 64GB of LPDDR5X at 8533 MT/s, with 136.5 GB/s of bandwidth and up to 59GB that can be allocated as VRAM. CNX Software describes the 64GB LPDDR5X-8533 configuration with a 1TB NVMe SSD preinstalled. Technetbook also reports the $2,599 pre-order price for the 1TB SSD configuration. Technetbook describes the memory as soldered, so plan on choosing the capacity you need at purchase.

LattePanda gives two performance figures, and they come from different tests:

  • Small-model comparison (product page): LattePanda's own test used OpenVINO GenAI 2026.4 and Qwen3.5-4B-int4 on the iGPU. The Sigma 2 reached 32 tokens/s at 38 W. The original Sigma reached 17 tokens/s at 35 W.
  • Large MoE model (press release): Qwen3.6-35B-A3B at "about 30 tokens/s" using INT4 and OpenVINO. CNX Software, citing LattePanda's documentation, says OpenVINO Model Server (OVMS) serves this model through an OpenAI-compatible API on port 8080, and repeats the roughly 30 tokens/s figure as LattePanda's claim.

Both numbers are the maker's and neither is independently verified. The 35B-A3B model is a mixture-of-experts design with roughly 3B parameters active per token. That helps explain how a 35B-parameter model can run at that speed on an iGPU, so don't expect similar throughput from a dense model of the same size.

I/O: built like a small server

The interface list explains the "server" label. From LattePanda's product page:

  • 3x M.2 M-Key, sharing a combined 3.3 V, 8 A power budget
  • 1x M.2 E-Key
  • OCuLink, PCIe 4.0 x2
  • 2x Thunderbolt 4
  • 2x 10GbE
  • RP2350 MCU co-processor for GPIO

CNX Software adds lane-level detail. The three M-Key slots are PCIe 4.0 x1, PCIe 4.0 x4 and PCIe 5.0 x2. The E-Key slot carries PCIe 4.0 x1, CNVio3 and USB 2.0. The two 10GbE ports use Realtek RTL8127 controllers. For display, CNX lists HDMI 2.1 up to 4096 x 2304 at 60 Hz, DisplayPort 2.1 over USB-C up to 7680 x 4320 at 60 Hz, and eDP 1.5 up to 3840 x 2400 at 120 Hz.

The three M-Key slots are not equal. Only one is x4, and one is a single lane, so put the main NVMe drive in the x4 slot. The shared 3.3 V, 8 A budget also matters if you plan to fill all three slots with high-power SSDs or accelerator cards. LattePanda's specification table also lists two USB 3.2 Gen 2 Type-A ports (10Gbps) and two USB 2.0 Type-A ports (480Mbps). Technetbook additionally mentions RS-232/RS-485 pins and a built-in turbine blower cooler, which we could not confirm on LattePanda's page or in the release, so treat those as secondhand.

Power comes in through USB-C PD at 20 V or a 5.5 x 2.5 mm barrel jack that accepts 15 to 20 V DC, per the product page.

The RP2350 side

As on earlier LattePanda boards, GPIO is handled by a separate microcontroller instead of the x86 SoC. The Sigma 2 uses Raspberry Pi's RP2350. LattePanda's release says it ships with MicroPython firmware preloaded, and Arduino development is also possible. CNX Software describes a 30+4-pin GPIO header driven by the RP2350. This arrangement puts real-time I/O such as sensor polling, PWM and bit-banged protocols on a microcontroller with deterministic timing, while the x86 side runs inference and networking. It also means the I/O code runs as firmware on the RP2350 rather than as ordinary Linux code on the x86 side.

Software

LattePanda lists Windows 11 and Ubuntu 24.04 or later as supported operating systems. The product page says Ubuntu, drivers, OVMS and OpenClaw come preinstalled on the built-in SSD. CNX Software, citing LattePanda's documentation, says the preinstalled image is Ubuntu 26.04. That fits within "24.04+", but check the image version on arrival if your toolchain depends on a particular release.

Dimensions: 146 x 102 mm

LattePanda's product page and press release both give 146 x 102 mm. Technetbook's summary prints a different and larger figure. We're using the maker's dimension.

What It Means for Makers

The Sigma 2 is not a hobby SBC, and its price is that of a small workstation. It is for people building fixed installations that need a local LLM, fast networking and some real-world I/O on one board: on-premises assistants, vision or inspection cells, edge gateways that must keep data local, robotics controllers that need a large model next to the motor and sensor code, and lab or classroom demo machines. In that role it can replace a mini PC plus a separate microcontroller board plus a 10GbE NIC. LattePanda's product page pitches the OCuLink port for connecting an external graphics card, which leaves room for an external GPU or other PCIe device if the integrated Arc B390 falls short.

The value depends on how you count. If you need the 64GB of fast unified memory, dual 10GbE and three NVMe slots, this is the configuration the board is built for. If your model fits easily in 16GB, or you don't need 10GbE, a cheaper mini PC paired with a separate microcontroller board is likely to cover the same work for much less. Treat the 180 TOPS headline as a sum across three engines that LattePanda says is not combined for a single workload, and base your sizing on the GPU's 122 TOPS INT8 and the 136.5 GB/s memory bandwidth.

Status and pricing: Announced October 22, 2026, and on pre-sale now. The pre-sale price is $2,599 USD through mid-November, according to LattePanda's release, after which it rises to $2,799 USD. Technetbook reports the $2,599 price for the 1TB SSD configuration. None of the sources give a ship date.

Sources