Specifications
Product
Firefly AIBOX-K3 Edge AI mini PC
SoC
SpacemiT Key Stone K3
CPU
8× 64-bit RISC-V X100 “big” cores up to 2.4 GHz, RVA23 compliant
CPU Performance
130 KDMIPS (comparable to Rockchip RK3588)
AI Cores
8× RISC-V A100 AI cores, 1024-bit RVV1.0 vector datapath
AI Performance
Up to 60 TOPS INT4 (sparse) with TCM + DMA acceleration channels
Model Support
Large models up to 30B parameters including Mixture-of-Experts (MoE)
GPU
Imagination BXM4-64-MC1 — Vulkan 1.3, OpenCL 3.0, OpenGL ES 1.1/2.0/3.2
Video Decode
H.265 / H.264 / VP9 up to 4K @ 120 fps
Video Encode
H.265 / H.264 up to 4K @ 60 fps
Memory
8 GB / 16 GB / 32 GB LPDDR5-6400 (51 GB/s)
On-board Storage
128 GB / 256 GB / 512 GB UFS 2.2
NVMe
M.2 socket (PCIe 3.0 ×4), 2242 or 2280
Display
HDMI 2.0 up to 4K @ 60 Hz
Networking
Dual Gigabit Ethernet RJ45
USB
2× USB 3.0 Type-A (1 A), 1× USB 3.0 DRD Type-C
Console
USB Type-C serial debug console
Power Input
9–20 V DC via 5.5 × 2.1 mm barrel jack (12 V / 5 A recommended)
Operating Temperature
−20 °C to +60 °C (storage −20 to +70 °C)
Humidity
10% to 90% RH non-condensing
Dimensions
93.4 × 93.4 × 50.0 mm
Weight
Approximately 500 g
Enclosure
Industrial-grade aluminium with porous hexagonal structure and strip grille ventilation
Indicators
Power button, boot button, status LED
Overview
The Firefly AIBOX-K3 is an Edge AI mini PC built on the SpacemiT Key Stone K3, an octa-core 64-bit RISC-V SoC running the X100 “big” cores at up to 2.4 GHz alongside eight dedicated A100 AI cores. Firefly rates the integrated AI engine at up to 60 TOPS INT4, which is enough to run locally hosted large language models up to roughly 30 billion parameters, including Mixture-of-Experts architectures, without sending data off the box.
What makes it interesting is that it is a RISC-V platform with a desktop-class software story rather than an MCU-class one. RVA23 compliance on the CPU, a 1024-bit RVV 1.0 vector datapath in the AI cores, and an Imagination BXM4-64-MC1 GPU with Vulkan 1.3 and OpenCL 3.0 mean the board can host a full accelerator runtime — Firefly supports multimodal acceleration on it — alongside conventional Linux services.
The 93.4 × 93.4 × 50 mm aluminium chassis carries dual Gigabit Ethernet, three USB 3.0 ports, HDMI 2.0 at 4K60, and an M.2 PCIe 3.0 ×4 socket that takes a 2242 or 2280 NVMe SSD on top of the 128–512 GB of UFS 2.2. Memory runs to 32 GB LPDDR5-6400 at roughly 51 GB/s, which is the bandwidth that actually determines LLM token throughput. Operating range is −20 to +60 °C on a 9–20 V DC barrel input.
Key Benefits
60 TOPS on a RISC-V host. Local inference up to 30B-parameter and MoE models without a cloud round-trip. Real memory bandwidth. 32 GB LPDDR5-6400 at ~51 GB/s, which is what decides token throughput rather than peak TOPS. Open accelerator stack. RVA23 CPU, 1024-bit RVV 1.0 vector units and a Vulkan 1.3 / OpenCL 3.0 GPU — not a vendor-locked NPU. Deployable chassis. Fan-vented aluminium industrial enclosure, dual GbE, M.2 NVMe expansion and a −20 °C cold start.
Applications
On-premises LLM and multimodal inference, private AI assistants and document processing, industrial visual inspection, robotics and autonomous platforms, smart-city video analytics, research and university RISC-V workloads, air-gapped data environments.
Request a Quote — FIREFLY AIBOX-K3 — SPACEMIT K3 RISC-V EDGE AI MINI PC, 60 TOPS
QS Compute — global B2B supply of AI computing hardware, edge AI systems and accelerators. Volume pricing, 15-day sample lead time.
Get Your Quote →