Maybe you missed it? Orange Pi 6 Plus Review.
You are currently reading:
An overview of SpacemiT’s latest K3 RISC-V boards built for running LLMs, along with a hands-on review of the new K3 Pico-ITX Mini PC, capable of running Ubuntu Mainline and various other operating systems.
Part I: An Introduction to the SpacemiT K3 RISC-V Series Boards
SpacemiT has rolled out an exciting new range of products built around the K3 ecosystem, delivering a great balance of affordability and power efficiency.
In recent years, China has made remarkable strides in the chip industry, fueled by its adoption of open-source software solutions and, more recently, the RISC-V architecture. SpacemiT Inc. is one of the leading Chinese semiconductor companies, specializing in developing next generation, high performance AI CPUs based on the RISCV architecture. Their mission is to create the best native computing platform for the new AI era, with a focus on large model AI workloads.
This article dives into the new K3 RISC-V series boards, designed to tackle massive models with support for up to 30 billion parameters. In the second part, we’ll put the K3 Pico-ITX Mini PC to the test.
Why “capable of running 30B parameter LLMs” is a big deal?
Running a 30-billion-parameter model on your own machine is no easy task. Typically, this requires:
- High memory bandwidth
- Efficient parallel compute
- Optimized AI acceleration
- Support for quantization and model compression
- A robust software stack (e.g., ONNX Runtime)
1️⃣SpacemiT K3 high-performance RISC-V processor
The K3 is a cutting-edge, high-performance RISC-V CPU built for advanced AI applications.
The new SpacemiT K3 chip series uses RISC-V homogeneous integrated computing technology, packing 8 high-performance X100 cores and 8 ultra-wide parallel AI A100 cores developed by SpacemiT. Together, they deliver 130 KDMIPS of general computing power and 60 TOPS of AI performance, capable of running 30-billion-parameter models with ease.
he K3 is the more advanced predecessor to the K1, featuring 60 TOPS of computing power and supporting double the RAM capacity compared to the previous K1 chip. It can handle up to 32GB of DDR5 RAM and is capable of running large AI models.
SpacemiT K3 CPU Block Diagram
Target market
The K3 chip series is designed for AI consumer hardware, including smart home gadgets, AI-powered conference and office tools, AI content creation devices, AI-driven e-commerce setups, retail systems, and users who want to run local LLMs, among other applications. All of this offers real value to developers, as mainstream Linux distributions, toolchains, and runtimes work smoothly right out of the box without any tweaks.
Main Highlights
| Feature | Required by RVA23 | Supported by SpacemiT K3 |
| RV64I base | ✔ | ✔ |
| M/A/F/D/C | ✔ | ✔ |
| Bitmanip (B) | ✔ | ✔ |
| Vector (V / RVV 1.0) | ✔ | ✔ (1024bit) |
| Hypervisor (H) | ✔ (Sprofile) | ✔ |
| AIA / IOMMU | Optional | ✔ |
| Full RVA23 compliance | ✔ | ✔ |
Completely compliant with the RVA23 profile.
What is RVA23 and why is it important?
RVA23 is a “standard package” of features that every modern 64 bit RISC-V processor must have if it wants to run big operating systems like Linux or Android smoothly. Think of it like a minimum spec list for RISC-V CPUs. Without RVA23, every chip maker could pick random features, and software would break.
With RVA23, everyone agrees on the same baseline.
What’s inside RVA23?
RVA23 requires a CPU to support:
- Basic math + advanced math (multiplication, division, floating point)
- Compressed instructions (smaller, faster code)
- Atomic operations (needed for multithreading)
- Bit manipulation (faster crypto, hashing, graphics)
- Vector instructions (the big one)
- Performance counters (profiling, debugging)
- Memory model guarantees (so OSes behave predictably)
- Optional: hypervisor support (for virtualization, in the Sprofile)
Why developers care?
RVA23 means:
- No more guessing which extensions a CPU supports
- One binary target for Linux/Android
- Better performance guarantees
- A stable foundation for future RISC V software ecosystems
It’s the difference between “maybe it works” and “it definitely works
In one sentence
RVA23 is the official checklist that ensures all modern 64 bit RISC V CPUs have the same essential features — especially vectors — so software runs reliably everywhere.
Here’s a closer look at the full range of capabilities offered by the SpacemiT K3:
2️⃣K3 Pico-ITX single board computer (Also known as Milk-V Jupiter 2)
The K3 Pico-ITX is a compact single-board computer packing up to 60 TOPS of AI performance. It comes with a unified memory setup, 8 CPU cores, and 8 AI acceleration cores, plus fast onboard UFS storage and a 10Gb optical network port. Its design boosts computing power and efficiency, making it great for tasks like scientific computing and AI.
The K3 Pico-ITX, designed in a compact 2.5-inch Pico-ITX Plus form factor, is perfect for tight spaces. It features dual M.2 expansion slots and offers interfaces for real-time motion control and system management.
Packed with versatile I/O options and built with an industrial-grade design, the K3 Pico-ITX makes it easy to evaluate and integrate systems quickly, helping speed up time to market.
SpacemiT K3 Pico-ITX Block Diagram
Specifications
| Module | Description |
| SoC | SpacemiT K3 |
| Processor | SpacemiT K3, 8 cores @2.4 GHz, 60 TOPS AI performance, RVA23 compliant, supports IME vector extensions and full virtualization |
| GPU | IMG BXM-4-64-MC1 OpenGL ES1.1 / ES2.0 / ES3.2 OpenCL 3.0 Vulkan 1.3 |
| Display | DP Type-C: up to 4K (3840 × 2160) @ 60 Hz 40-pin eDP: up to 2.5K (2560 × 1600) @ 90 Hz |
| Memory | Dual-channel 2 × 32-bit LPDDR5 6400 MT/s, 16 GB / 32 GB options |
| Local Storage | UFS 2.2, 128 GB / 256 GB options |
| Storage Expansion | M.2 M-Key (PCIe Gen3 ×4), supports 2280 NVMe SSD |
| HS Expansion | M.2 B-Key (PCIe Gen3 ×2 + USB), supports 2242/3042 cards |
| Real-Time Expansion | FPC connector for EtherCAT, 5 × CAN-FD, SPI, I²C, UART, etc. |
| Wireless | Onboard Wi-Fi 6 + BT 5.2, dual-band, dual-antenna, 802.11 a/b/g/n/ac/ax compliant |
| Wired Network | 1 × RJ45 Ethernet, 10/100/1000 Mbps adaptive |
| Optical Network | 10GbE SFP+ port, supports 10GBASE-R / 10GBASE-X, QinQ, MSI-X, WOL, and clustering |
| Audio | Onboard CODEC, internal audio input/output |
| USB | 2 × USB 3.2 Gen1 Type-C (1 full-featured, one OTG) 4 × USB 2.0 Type-A Host |
| Debug | UART, JTAG, and 3 side buttons for power, reset, and firmware update |
| System Management | Onboard EC controller for power, thermal, and system status; includes I²C/UART/GPIO expansion |
| Form Factor | 100 × 86 mm, Pico-ITX Plus single-board computer, approx. size of a 2.5″ drive |
| OS | Pre-installed Bianbu 3.0; supports Ubuntu 26.04, OpenHarmony 6.0, OpenKylin, Deepin, Fedora, etc. |
| Power Input | Dual Type-C USB-PD (65 W) or ATX 2-pin 12 V @ 7 A |
| Reliability | ESD protection: board: ±4 kV (contact), ±8 kV (air); system: ±6 kV (contact), ±12 kV (air) Compliant with CCC, CE, and FCC; operating temp: -20 °C ～ 70 °C (consumer) / -40 °C ～ 85 °C (industrial) |
| Clock | Onboard RTC with battery interface, supports G3 state |
| Structure | Optional single-board or fan-cooled heatsink assembly Optional single-board or fully-metal industrial chassis Optional real-time board, touchscreen, terminal blocks |
| Mainline software support | ✅ Ubuntu, OpenWrt ✅ OpenHarmony ✅ OpenKylin ✅ Bianbu |
| Link to the product | Milk-V Jupiter2 | Your All-in-One Development Platform |
Note: The M.2 B-Key PCIe ×2 lanes share bandwidth with the M-Key slot; if both are in use, the M.2 M-Key runs at PCIe Gen3 ×2.
Optional Components
| Category | Name | Description | Interface |
| Peripheral | SSD | 980 NVMe™ M.2 SSD, PCIe Gen 3.0 ×4, NVMe 1.4, sequential read up to 3,500 MB/s, sequential write up to 3,000 MB/s | M.2 M-Key 2280 |
| SSD | B+M NVMe SSD 128 GB | M.2 B-Key 2242 | |
| 4G Module | EM05 | M.2 B-Key 3042 | |
| SATA Expansion Card | PCIe to 5 × SATA interface | M.2 M-Key 2280 | |
| Docking Station | Type-C dock for HD 4K display, PD charging, USB 3.0 for tablets/laptops (3-in-1) | Full-featured Type-C | |
| Ultra HD Display | 16″ 2.5K LCD, 90 Hz, 2560 × 1600 | 40-pin eDP | |
| Real-time Control Expansion Board | 19 V input, 5 × CAN-FD, EtherCAT, RS232 & RS485 I/O, industrial-grade isolation protectiosn | FPC | |
| Structural Accessory | Embedded Fan Heatsink | Custom aluminum design | FAN |
| Metal Chassis | 120 × 120 × 48 mm self-developed metal chassis | BTN |
3️⃣SpacemiT K3-CoM260 Developer Kit
A Full-Stack RISC-V Robotics Development Kit
The SpacemiT K3-CoM260 Developer Kit packs an 8-core general-purpose CPU and an 8-core AI CPU into a compact compute module with a reference carrier board. It offers up to 60 TOPS of AI performance and 130 KDMIPS of general computing power, making it capable of running 30B large models smoothly.
Powered by the RISC-V architecture, the kit makes it easy to run popular AI models while sticking to standard CPU programming methods, enabling smooth, cost-free migration of AI algorithms. It’s hardware compatibility with the Orin Nano makes it a complete solution for edge AI development and testing, perfect for AI appliances, service robots, and autonomous edge devices.
SpacemiT K3 CoM260 (CASE + Board)
Key Features
| Feature Category | Description |
| High Performance Edge AI | Up to 60 TOPS AI compute, enabling smooth 30B parameter model inference and multi stream concurrent workloads |
| CPU + AI CPU Architecture | Codesigned general purpose CPU + AI CPU for balanced system compute and high performance AI inference |
| Modular Design | Compute module + carrier board form factor; fully compatible with K3 CoM260 series for easy integration |
| Developer Friendly | Standard CPU programming model allows zero cost migration of AI algorithms and system software |
| Advanced RISC-V Architecture | First compact RISC V edge platform supporting the RVA23 profile |
| Rich Standard Interfaces | Comprehensive high speed I/O for cameras, sensors, and peripherals |
| Built for Real World Applications | Ideal for AI appliances, service robots, and autonomous edge agents |
| Engineering Ready Support | Supports custom software, carrier board design, and full system integration to accelerate deployment |
Extensive software compatibility
The K3 now has native support in Linux kernel 7.0, and work is underway to upstream OpenSBI, U-Boot, LLVM, and other core software. For more information, see the details on the project’s GitHub wiki. (https://github.com/spacemit-com/linux/wiki).
OS compatibility now includes Ubuntu 26.04 with support for RISC-V architecture.
K3 currently supports a range of open-source operating systems, including Ubuntu, OpenHarmony, OpenKylin, and OpenEuler. Additional operating systems and widely used tools are in the process of being adapted to expand compatibility.
Official mainline Ubuntu support.
February 5, 2026 – Canonical, the company behind Ubuntu and other open-source solutions, has teamed up with SpacemiT to bring the Ubuntu operating system to the K3/K1 RISC-V AI CPU platform. This partnership highlights the deep integration of open-source operating systems with open RISC-V chips, delivering powerful, flexible, and reliable computing options for developers around the globe.
Continuously updated download links can be found at:
(https://www.spacemit.com/community/eco-software?id=spacemit-open-source-soft)
Extra support through official partner channels
As a hardware partner of SpacemiT, Milk-V provides user support and shares technical information through its community forum.
The Banana Pi team also provides software and hardware support through their community forum.
What about official OpenWrt mainline support?
The answer is affirmative. According to company sources, it’s currently in development. The official source tree and Firmware updates for SpacemiT’s latest boards, including the K3, will be available and easily accessible through OpenWrt official website.
Supporting the Spine Triton kernel development system
Triton, developed by OpenAI, is a high-level language that lets developers create custom AI kernels with impressive speed. In simpler terms, it’s an open-source framework and library for writing highly efficient GPU code, designed to boost productivity beyond what CUDA offers.
Spine-Triton is a compiler and kernel-writing framework supports SpacemiT RISC-V chips like the K1 and K3, making it easy to create high-performance kernels—like MatMul, Softmax, and FlashAttention—using Triton-style code.
Why does SpacemiT care about Triton?
If SpacemiT gets Triton running efficiently on their RISCV AI CPU, it means:
- Developers can write GPU style kernels for K3
- AI operators can be optimized without handcoding assembly
- RISCV becomes more competitive with NVIDIA/ARM
- K3 becomes a real platform for custom AI workloads
This is a smart strategy to create a modern AI software ecosystem that complements their hardware.
Although Triton has many advantages, it has limitations when targeting CPUs. SpacemiT aims for something bigger: enabling Triton written operators to run efficiently on RISC-V AI CPUs.
- Deployment info: https://github.com/spacemit-com/spine-triton
- Example: https://forum.spacemit.com/t/topic/872
Hardware designed to work with AI agents
The SpacemiT K3 is a RISCV AI processor optimized for on device AI agents, integrating 8 general purpose X100 CPU cores with 8 A100 AI accelerated cores featuring 1024bit vector units and delivering up to 60 TOPS of performance. With both CPU and AI cores sharing a unified architecture and running a single Linux OS, it eliminates the overhead of separate NPUs and drivers, making it well suited for real time LLM inference, autonomous decision loops, and robotics oriented agent workloads. Boards like the Banana Pi SM10, based on the K3, are marketed for edge AI agent development and can run 30B parameter models locally at practical speeds, underscoring the platform’s design for agent style autonomy.
The K3’s design makes it a great fit for local AI autonomy, with support for full Linux distributions like Ubuntu, OpenHarmony, and OpenKylin, plus fast LPDDR5 6400 memory for handling large models. Its 15–35W power range works well for robots, kiosks, and smart devices. The shared RISC V ISA across compute and AI cores lets agent logic, planning, and neural inference run in the same memory space with minimal latency—just what modern agent frameworks call for. In real world terms, the K3 stands out as one of the first serious RISC V platforms capable of running on device LLM agents, driving applications from service robots to industrial automation and sensor driven autonomous systems.
Platforms that are already compatible with the K3 architecture.
| Platform | Description |
| Zenow | On-device AI knowledge assistant desktop app. All data processing stays local, ensuring privacy. Supports multi-model management, intelligent chat, knowledge base Q&A, voice interaction, etc. |
| Yumeet | An on-device AI meeting assistant desktop application offering local audio processing, transcription, translation, and summarization to enhance meeting efficiency while safeguarding sensitive data. |
| Seewise | multi-modal vision processing: image search by text, video search by text. |
RISC V AI Acceleration in 2026: How SpacemiT K3 Challenges Jetson at the Edge
SpacemiT K3 vs NVIDIA Jetson Series – Performance Comparison
Here’s a clear, side by side comparison table highlighting the performance positioning of the SpacemiT K3, NVIDIA Jetson Orin Nano, and NVIDIA Jetson AGX Orin.
| Category | SpacemiT K3 | NVIDIA Jetson Orin Nano | NVIDIA Jetson AGX Orin |
| AI Compute (TOPS) | 60 TOPS | 20 TOPS | 275 TOPS |
| LLM Capability | Smooth 30B model inference | Suitable for small models (≤7B) | Can run 70B+ models (quantized) |
| AI Architecture | 8× A100 AI cores (1024 bit RVV vector engines) | 32 Tensor Cores | 2048 CUDA cores + 64 Tensor Cores |
| CPU Architecture | 8× X100 RISC V @ 2.4GHz | 6× ARM Cortex A78AE | 12× ARM Cortex A78AE |
| Strengths | Efficient LLM inference, strong vector AI, low power | Lowest cost, good for vision tasks | Maximum performance for robotics & multimodel AI |
| Weaknesses | No CUDA ecosystem; new platform | Limited AI compute | High cost & high power draw |
| Power (TDP) | 15–25W | 7–15W | 15–60W |
| Best Use Cases | Edge LLMs, AI appliances, RISC V development | Lightweight robotics, cameras, CV | Heavy robotics, multi camera AI, high end edge compute |
| Estimate price | ~$300 | $280 – $580 USD | $1,300–$1,900 USD |
