Sipeed K3 Boards Bring Compact RISC V AI Power With Up to 32 GB LPDDR5 and Local 30B LLM Support
Sipeed has introduced its new K3 series single board computers built around the SpacemiT Key Stone K3, giving the edge AI and embedded market a fresh RISC V option that is clearly targeting local inference, networking, and compact AI compute workloads. The pitch is ambitious for the size: up to 60 TOPS of AI performance, support for models up to 30B parameters, and board designs that land in compact developer friendly formats including a Pico ITX class layout and a CoM260 module compatible with Jetson Orin Nano carrier ecosystems.
At the silicon level, the K3 platform combines 8 X100 RISC V CPU cores running up to 2.4 GHz with 8 A100 AI oriented compute cores, and SpacemiT positions the chip at up to 130 KDMIPS alongside up to 60 TOPS of AI throughput. SpacemiT also says the platform can run local language models up to 30B parameters at more than 10 tokens per second, while Sipeed’s own showcase claims up to 15 tokens per second with Qwen 3.5 35B in a demonstrated workload. That makes the K3 a serious attempt to push local LLM inference into a much smaller and more open hardware footprint than many buyers may expect from current AI edge systems.
The board side is where this launch gets especially interesting. The K3 CoM260 Developer Kit uses a 69.6 mm by 45 mm compute module and is designed for Jetson Orin Nano carrier compatibility, which could lower migration friction for developers already invested in that ecosystem. The K3 Pico ITX board comes in at roughly 100 mm by 86 mm and adds a richer edge oriented I O mix that includes 10 GbE, Gigabit LAN, dual USB Type C with power delivery and display support, M.2 expansion, and onboard LPDDR5 options up to 32 GB depending on SKU. Sipeed also positions the software side around Ubuntu 26.04 and ROS support, which is a practical move if it wants the platform to be taken seriously for robotics, vision, and local inference development.
Core specifications are summarized below from the currently published K3 platform details.
| Hardcore Specs | Details |
|---|---|
| SoC | SpacemiT Key Stone K3 (8x X100 + 8x A100 RISC-V Cores) |
| AI Performance | 60 TOPS (INT4), Supports BF16/FP16/INT4 |
| Memory (RAM) | 8/16/32GB LPDDR5 @ 6400MT/s (51GB/s Bandwidth) |
| Storage | Supports eMMC 5.1, SD Card, M.2 NVMe SSD (PCIe Gen3 x4) |
| Multimedia | Decode: 4K@120fps (H.265/VP9) | Encode: 4K@60fps |
| OS Support | Bianbu OS (Debian based), Docker Support, RISC-V KVM Virtualization |
| Dimensions | 103mm x 90.5mm x 35mm (with heatsink) |
Pricing is also aggressive for a niche RISC V AI platform, though it is not exactly low cost. Current listing information places the 8 GB K3 boards around 299.00$ to 309.00$, while 32 GB configurations rise to roughly 629.00$ to 639.00$. That puts them above NVIDIA’s Jetson Orin Nano Super Developer Kit on Amazon, where the 8 GB kit is listed at 249.00$ with a 6 core ARM Cortex A78AE CPU. The tradeoff is that Sipeed is offering a very different value proposition built around RISC V, larger memory options, and a stronger local LLM angle rather than a direct like for like Jetson clone.
For the wider market, the K3 series matters less as a direct threat to NVIDIA and more as a sign that compact AI hardware is beginning to diversify in a meaningful way. If Sipeed and SpacemiT can back these specifications with strong software maturity, stable developer tools, and real world local inference performance, the K3 family could become one of the more compelling RISC V edge AI launches of 2026 for builders who want something more open and more memory rich than entry level ARM alternatives.
Would you take a compact RISC V AI board like this over a Jetson class system for local LLM testing and edge projects?
