What happened?
The Sipeed MaixStation DGX Spark arrives as the new Chinese rival. Sipeed announced the MaixStation, a compact local AI computer based on the Axera AX8860P chip. The system offers 128 GB of LPDDR5X memory with 273 GB/s bandwidth, a 256 TOPS native FP4 NPU, and around 40 W power consumption. The company claims better performance than NVIDIA's DGX Spark on models like Qwen 27B, with up to 25% faster prefill and 30% faster decode, at an estimated $3,999 for the 128 GB version — roughly half the NVIDIA competitor's price.
The announcement comes amid high demand for local AI hardware, with Sipeed positioning the product as an accessible and energy-efficient alternative. 32 GB versions and a 512 GB Pro are also planned, with a Kickstarter launch expected by the end of 2026.
Why does it matter?
The local AI station market is growing rapidly, driven by the need to run large models without relying on the cloud. NVIDIA's DGX Spark is a reference, but its high price and energy use limit adoption. Chinese solutions like the MaixStation could democratize access to high-performance local inference.
The Axera AX8860 uses a 6 nm process, 16 Cortex-A55 cores, and supports frameworks like vLLM, SGLang, and PyTorch. Its compact architecture (about 363 cm³) allows portable use, even with a power bank.
What changes in practice?
Developers and AI enthusiasts can expect a cheaper option to run 27B to 125B models locally, with support for agents, video generation, and simple clusters via cable without expensive switches. Performance claims are from Sipeed and await independent benchmarks.
Source: Sipeed announcements and coverage from Tom's Hardware and IT Home. Official site.
Image credit: Sipeed — Source: X/@SipeedLab
By GeekikiBot
