The SBC & computing space is growing faster than ever with the YuzukiNeko, a Raspberry Pi Pico-sized RISC-V board with as little as 8MB of RAM that can still run Linux. As well as new compact workstations from Minisforum and ASUS packing as much as 192GB of unified memory and hardware built specifically for demanding local AI workloads.
Despite the massive gap in capability, all three have something in common: they're squeezing increasingly serious computing platforms into surprisingly small packages.
YuzukiNeko puts RISC-V Linux in a Raspberry Pi Pico-sized board
YuzukiHD's new YuzukiNeko might be the most interesting little board we've seen for a while.
At roughly 51 × 21mm, YuzukiNeko follows the familiar “gum stick” sbc form factor with other similarly sized boards like; Luckfox Lyra, Milk-V Duo, & Ox64. Difference is the Neko is using an Allwinner F101 multimedia SoC built around a T-Head XuanTie C907 RISC-V CPU.
The C907 is a 64-bit RISC-V core clocked at up to 1.008GHz, with 32KB of instruction cache and 16KB of data cache. The F101 comes in two memory configurations. The F101S2 pairs the processor with 8MB of 16-bit PSRAM, while the F101S3, the variant used here, steps up to 16MB. That's enough to push YuzukiNeko beyond typical microcontroller territory and into lightweight Linux use.
The board itself adds 16MB of external SPI NOR flash, a microSD slot, a USB-C port, an FEL button, and two rows of 20-pin GPIO headers arranged in the familiar Pico-style layout.
And yes, it can run Linux. Support is based on the mainline Linux 7.2 kernel alongside Buildroot, while Zephyr RTOS is available as a lighter alternative for projects that don't need a full Linux environment.
That puts YuzukiNeko in an awesome spot between a microcontroller development board and a conventional SBC.
There's a surprising amount packed into the F101
The processor is arguably even more interesting than the board's tiny size. Allwinner designed the F101 for multimedia applications.
The chip supports RGB888 output at up to 1280 × 720 at 60fps, single-link LVDS up to 1366 × 768 at 60fps, and 4-lane MIPI DSI up to 1280 × 800 at 60fps. There's also a G2D hardware accelerator for rotation and image mixing, JPEG decoding up to 16384 × 16384, and hardware JPEG/MJPEG encoding up to 720p30.
Beyond that, the F101 provides I2S/PCM and OWA digital audio, an integrated DAC with analog headphone output, a USB 2.0 DRD port, SDIO 2.0/3.0, two SPI controllers, six UARTs, three I2C/TWI interfaces, four-channel PWM, infrared receive, and both general-purpose and touch-panel ADC channels.
Not every function can be used at the same time since many share the same physical pins, but that's still a considerable amount of I/O packed into such a small chip.
It's completely open hardware
YuzukiNeko isn't just another development board announcement.
The hardware is released under the CC0 1.0 public-domain dedication, and YuzukiHD has published the full design resources on GitHub, including schematics, the EasyEDA PCB layout, Gerber files, a 3D STEP model, and a bill of materials. (The Github link is provided below for anyone who wants to check it out.)
That makes YuzukiNeko a solid starting point for anyone wanting to build custom embedded devices around the F101, rather than simply using it as another board to drop onto a breadboard.
Unfortunately it’s not currently available for purchase yet.
YuzukiNeko isn't for sale as of this writing, although YuzukiHD's previous boards, including the Avaota A1 and Yuzuki Chameleon, have eventually made their way to online retailers. For some context on where pricing could land, one forum post puts the bare F101 chip at around 18 yuan, or roughly $2.70. That suggests a finished YuzukiNeko could plausibly come in under $10 once it becomes available.
Even without confirmed pricing, the combination of RISC-V, Linux support, real display and multimedia hardware, open hardware licensing, and a Pico-sized footprint makes this one worth watching.
Blu Dot surpasses 2,000% ROAS with self-serve CTV ads
Blu Dot used Roku Ads Manager to drive incredible results for its furniture sales event. Its strategy hinged on custom audiences and retargeting, where intent was strongest.
“Roku has been a top performer,” said Blu Dot’s Claire Folkestad. “We have seen…CPMs lower than any other CTV partner we've worked with.”
Minisforum goes in the opposite direction with 192GB of memory
Minisforum's new MS-S1 MAX-P495 shows just how much computing power can now fit into a mini workstation.
Introduced at IFA 2026 in Berlin, the MS-S1 MAX-P495 is built around AMD's Ryzen AI Max+ PRO 495, the "Gorgon Halo" chip. It combines 16 Zen 5 CPU cores and 32 threads, boosting up to 5.2GHz, with Radeon 8065S integrated graphics and dedicated NPU acceleration. Minisforum rates the complete platform at up to 131 TOPS of AI compute.
But the standout specification is memory.
The MS-S1 MAX-P495 supports up to 192GB of LPDDR5X-8533 unified memory, a 50 percent increase over the 128GB ceiling of the previous PRO 395 model. As much as 160GB of that memory can be allocated to graphics.
That's an enormous amount of GPU-accessible memory for a chassis measuring just 222.1 × 206.3 × 77.1mm.
Why 192GB matters for local AI
GPU memory remains one of the biggest constraints when experimenting with large AI models locally. A powerful desktop might have 64GB or 128GB of system RAM but only 12GB, 16GB, or 24GB of dedicated VRAM on its graphics card.
Unified memory changes that relationship by allowing the CPU and GPU to access the same large memory pool. With up to 160GB available to the graphics side, Minisforum's machine can take on workloads that might otherwise require professional GPUs or multiple graphics cards simply to fit a model into memory.
Minisforum is positioning the MS-S1 MAX-P495 for local inference, generative AI, computer vision, and other demanding edge-AI workloads. The company is also demonstrating the system in a cluster configuration. Two 192GB units linked together can reportedly run the 397-billion-parameter Qwen3.5-397B model locally at around 16 tokens per second.
It's worth keeping in mind that this is Minisforum's own benchmark rather than an independently verified result.
Pricing hasn't been officially confirmed, but Minisforum's teaser points toward roughly €7,000, with sales expected to begin in September 2026.
For something that still fits comfortably on a desk, that's a substantial amount of compute.
ASUS puts NVIDIA Blackwell and Grace into a tiny desktop
ASUS is approaching the same local AI market from a completely different direction.
The new ProArt GR1X is a compact desktop built around NVIDIA's RTX Spark platform.
RTX Spark pairs an NVIDIA Grace CPU with up to 20 cores with a Blackwell RTX GPU packing 6,144 CUDA cores and fifth-generation Tensor Cores, connected through NVLink-C2C. ASUS says the platform delivers up to 1 petaFLOP of FP4 AI performance and can be configured with up to 128GB of unified memory.
CUDA remains deeply established across AI development, rendering, and scientific computing. Pairing that ecosystem with Blackwell graphics and 128GB of unified memory makes the GR1X an interesting alternative to a traditional workstation packed with discrete GPUs.
ASUS says the GR1X can run LLMs with up to roughly 120 billion parameters, handle 3D scenes larger than 90GB, and generate 4K AI video locally.
And it's only 150mm wide
The hardware gets even more impressive when you consider the size of the mini PC.
ASUS says the ProArt GR1X measures just 150 × 150 × 51mm, roughly six inches square and two inches tall.
Despite the tiny chassis, ASUS designed it for sustained workloads rather than short bursts of performance. RTX Spark runs at up to 140W TDP in the GR1X, notably higher than the 80W ASUS is using for the same platform in its ProArt P16 and P14 laptops. A dedicated dual-fan cooling system helps keep that hardware under control.
Connectivity includes 10Gb Ethernet, Wi-Fi 7, Bluetooth 5.4, and support for up to four 4K displays.
ASUS is targeting creators and developers working with local AI agents, 3D rendering, and other GPU-heavy workloads. In other words, this isn't competing with the inexpensive N100 mini PCs that have flooded the market over the last few years. It's a tiny workstation.
Small computers are getting much more interesting
YuzukiNeko, the MS-S1 MAX-P495, and the ProArt GR1X couldn't be much more different. One has as little as 8MB of RAM, while another has up to 192GB. Yet all three point toward the same broader trend.
YuzukiNeko brings a 64-bit RISC-V CPU, Linux support, real display and multimedia hardware, and open hardware licensing to something the size of a Raspberry Pi Pico.
Minisforum is putting workstation-class AMD hardware and a massive unified memory pool into a compact desktop designed to run large AI models locally.
ASUS, meanwhile, is shrinking NVIDIA's Grace and Blackwell architecture into a 150mm mini PC with 128GB of unified memory and access to the CUDA ecosystem.
For makers, developers, homelab users, and anyone interested in doing more computing locally, that's a pretty interesting direction.
Small computers aren't just getting faster. They're becoming capable of workloads that would have required an entirely different class of hardware just a few years ago.