Sipeed’s MaixStation AI computer claims edge over Nvidia DGX Spark in Qwen workloads

MaixStation’s performance claims

Chinese electronics manufacturer Sipeed has announced its MaixStation AI computer, powered by Axera’s AX8860P processor, claiming it outperforms Nvidia’s DGX Spark and Apple’s Mac Mini in specific Qwen workloads. The AX8860P combines 16 Arm Cortex-A55 cores with AI acceleration capable of 256 TOPS, though Sipeed argues that practical inference performance matters more than theoretical metrics.

Sipeed’s 128GB MaixStation configuration reportedly slightly exceeds Nvidia’s DGX Spark in Qwen benchmarks, despite the latter being advertised at 1,000 sparse FP4 TOPS. Sipeed contends that real-world workloads cannot fully utilize Nvidia’s theoretical capacity, giving its hardware an edge in efficiency.

Power, pricing, and availability

The MaixStation has a typical power consumption of 40W, targeting low-power AI computing applications. The entry-level 32GB model is priced below $2,000, while the 128GB configuration is expected to cost under $4,000. For comparison, Nvidia’s DGX Spark reportedly sells for $6,949.99, up from its March 2025 pricing of $3,000–$4,000.

Sipeed plans to launch the MaixStation via a Kickstarter campaign, offering a $100 discount for early reservations made with a $30 deposit. The device measures 110 × 110 × 30 mm and includes USB 3.0 Type-C and a 10GbE SFP+ network interface.

The MaixStation supports large language models like Qwen3.8-27B on its 32GB model and Qwen3.8-Flash-Next 125B on its 128GB configuration. Multi-unit setups can handle even larger models, leveraging AXLink for inter-processor communication with approximately 80% efficiency.

Software and hardware compatibility

The AX8860P processor is manufactured using a 6nm process and offers 273GB/s memory bandwidth, matching the DGX Spark’s specifications. The MaixStation supports frameworks like PyTorch and ONNX, as well as OpenAI-compatible interfaces, making it adaptable for existing AI workflows.

What this means

Lazyfounder analysis — our interpretation, not reported fact.

Sipeed’s claims, if verified, could disrupt the edge AI hardware market by demonstrating that lower-power, lower-cost devices can compete with high-end alternatives in real-world workloads. This challenges the industry’s reliance on theoretical TOPS as the primary performance metric and may push competitors to prioritize practical benchmarks. For founders and operators, the MaixStation’s pricing and power efficiency could make it a viable option for deploying LLM-powered applications, particularly in cost-sensitive or power-constrained environments.

Key takeaways

• The MaixStation’s claimed performance challenges the assumption that higher theoretical TOPS always translate to better real-world AI inference, which could influence hardware procurement decisions for edge AI deployments.
• Its pricing and power efficiency may appeal to cost-sensitive founders and startups building LLM-powered applications, particularly in markets where affordability is critical.
• The shift toward practical benchmarks over theoretical metrics could pressure competitors like Nvidia to emphasize real-world workload performance in their marketing.

Related on Lazyfounder

• Nvidia
• Apple
• More technology news
• Modal shifts AI sandboxes to Nvidia’s Vera CPU, claims double throughput