Firefly's AIBOX-K3 Lands May 12 — A $349 RISC-V Edge AI Mini PC Running 30B Local LLMs at 60 TOPS

A Complete RISC-V Edge AI Mini PC for Local LLM Workloads Just Landed at $349

Firefly officially launched the AIBOX-K3 Edge AI mini PC on May 12, 2026, and it is the cleanest expression we have seen so far of how the SpacemiT K3 RISC-V platform translates into a finished desktop product. The AIBOX-K3 takes the SpacemiT K3 octa-core X100 RISC-V SoC — the same architecture showing up in the new wave of RISC-V single board computers — and packages it inside a fanless aluminum chassis with all the I/O, cooling, and power delivery a developer needs to run local AI workloads out of the box. Firefly is selling the entry configuration with 8GB LPDDR5 memory and 128GB UFS storage for $349, with a 32GB / 128GB SKU at the top of the range for $689.

For the local LLM community, embedded engineers building edge AI products, and developers comparing RISC-V mini PCs to Arm and x86 alternatives, the AIBOX-K3 is the kind of complete-system data point the platform has been waiting on. The SpacemiT K3 has been shipping in SBC and module form for several weeks; the AIBOX-K3 is the first widely available finished mini PC built around it.

The SpacemiT K3 SoC Inside the AIBOX-K3

The K3 is an octa-core RISC-V SoC built on SpacemiT's X100 high-performance application cores, fully compliant with the RVA23 profile that the RISC-V ecosystem has converged on for application-class workloads in 2026. Alongside the eight application cores the K3 integrates an AI engine rated at up to 60 TOPS, which is the headline number for edge inference workloads. SpacemiT reports the platform can sustain over 10 tokens per second on a 30-billion-parameter LLM running locally — the kind of throughput that makes interactive local-AI use cases practical on a fanless desktop.

Why RVA23 Compliance Matters for the Developer Ecosystem

The RVA23 profile compliance is the structural detail that determines whether software written for one RISC-V chip runs cleanly on another. Standardizing on RVA23 means the AIBOX-K3 inherits a software ecosystem that is converging quickly: Ubuntu 26.04 (developed in collaboration with Canonical), Bianbu OS 3.0, OpenHarmony, OpenKylin, Fedora, Deepin, and others. For developers evaluating RISC-V as a deployment target, broad RVA23 OS support is the prerequisite that makes the platform credible — and the AIBOX-K3 is shipping with it.

Security, Memory Protection, and Encryption Built In

One of the under-discussed details on the AIBOX-K3 is the security architecture the K3 SoC brings with it. The chip implements a three-level Machine/Supervisor/User privilege model, includes hardware mitigations against Spectre and Meltdown-class attacks, supports PMP and ePMP for physical memory protection, and integrates IOPMP for I/O-level memory protection. There are also built-in hardware encryption engines covering AES, SHA, RSA, and the SM2/SM3/SM4 cipher families.

Why That Matters for Edge AI Deployments

Edge AI workloads frequently process sensitive data — voice, video, medical imaging, financial transactions — and the integration of memory protection plus hardware encryption at the SoC level is what makes edge deployments defensible against the threat models customers actually care about. The AIBOX-K3 is not just a fast mini PC; it is a fast mini PC with the security primitives baked into the silicon, which is the configuration that turns it into a credible production platform rather than just a developer toy.

Local LLM Throughput on a Fanless Desktop Form Factor

The most striking practical claim Firefly is making is the over-10 tokens-per-second throughput on a 30B-parameter model. To put that in context, that is the lower edge of "conversationally usable" for many developers running interactive local LLM applications. It is achievable on a fanless desktop chassis with no discrete GPU, no liquid cooling, and no specialized accelerator card. For builders who have been waiting for the moment when local 30B models become practical on a $349 mini PC, this is that moment.

What the 60 TOPS NPU Enables Beyond LLMs

The 60 TOPS rating is sized for more than just LLM workloads. It covers the typical edge AI surface — vision pipelines, audio processing, speech recognition, on-device transcription, small multimodal models — and gives developers room to layer multiple inference workloads on the same box without saturating the accelerator. For developers building multi-modal applications on the edge, the AIBOX-K3 is a versatile compute platform rather than a single-purpose LLM appliance.

How the AIBOX-K3 Compares Against Arm and x86 Edge AI Mini PCs

The competitive read on the AIBOX-K3 is interesting. Against equivalent-class Arm edge AI mini PCs based on NVIDIA Jetson Orin Nano or Qualcomm-class SoCs, the K3 brings comparable AI throughput at a lower price point. Against x86 mini PCs based on Intel Wildcat Lake or AMD Ryzen AI parts, the K3 trades general-purpose CPU performance for lower thermal envelope and lower idle power. The RISC-V card is the cleaner answer for the deployments where deterministic, fanless, security-first edge AI is the priority, and where running mainstream Linux distributions with broad open-source tooling is acceptable.

The Storage and Memory Configurations

Firefly is offering the AIBOX-K3 in 8GB/128GB and 32GB/128GB configurations at $349 and $689 respectively. The 32GB configuration is the one that matters most for serious local LLM work — the headroom lets developers comfortably hold a 30B model in memory alongside the OS, application stack, and supporting tools. The 8GB entry point is sized for lighter edge AI workloads, vision pipelines, and embedded HMI applications.

The Setup Going Forward for the RISC-V Mini PC Market

For the RISC-V community, embedded developers, and local AI enthusiasts, the Firefly AIBOX-K3 is the kind of finished-product launch that converts the SpacemiT K3 platform from "interesting silicon" into "things you can buy and put on a desk today." The pricing is competitive, the software ecosystem is converging on RVA23 quickly, the security primitives are real, and the local LLM throughput claims land in the range that makes the box useful. The next watch items are independent third-party benchmarking and the availability ramp through Firefly's distribution channels. For builders comparing edge AI mini PC options in May 2026, the AIBOX-K3 deserves a spot on the shortlist.

Sources: CNX Software, May 12, 2026; LinuxGizmos, May 2026; Electronics Lab, May 2026; Firefly product page, May 2026.