What happened?
Chinese company Sipeed announced the MaixStation, a compact local AI system designed to compete with the NVIDIA DGX Spark and AMD Ryzen AI Halo. The device uses the Axera AX8860P chip, with 16 Arm Cortex-A55 cores and a proprietary NPU rated at 256 TOPS in native FP4 (no sparsity), manufactured on 6nm.
The 128GB configuration brings LPDDR5X on a 256-bit bus, delivering 273 GB/s bandwidth, identical to DGX Spark. Size is about 11 × 11 × 3 cm (363 cm³), half a Mac mini and a third of DGX Spark, with typical 40W consumption via USB-C PD, allowing use with a 65W power bank.
Why does this matter?
The expected price for the 128GB version is approximately $3,999, well below the roughly $6,950 of the DGX Spark 128GB and aligned with the Ryzen AI Halo 128GB. Sipeed claims superior performance in some metrics: on Qwen 27B (FP4, no speculative decoding), prefill about 10-25% faster and decode 10-30% faster than DGX Spark, though the numbers are still from internal tests and the chip is being tuned.
There are 32GB, 128GB and Pro 512GB versions (with multiple chips), plus MaixBlade PCIe card. Supports cascading via MCIO cable without expensive switches. Software includes Debian 13 with Linux 6.18, and support for vLLM, SGLang, PyTorch, llama.cpp and Ollama with NPU acceleration.
What changes in practice?
Reservations are open on Sipeed's site, with Kickstarter planned for the end of 2026 and shipping about three months after the campaign ends. The NPU toolchain is expected by the end of October. Preloaded models and OpenAI/Anthropic-compatible API facilitate immediate use for agents and local generation.
- Ideal for running dense 27B models on 32GB or larger MoE on 128GB.
- Portable and energy-efficient compared to traditional solutions.
- Focus on open stack and cable-based scalability.
Sources: Sipeed, Tom's Hardware, Igor's Lab.
Image credit: Sipeed — Source: Sipeed
By GeekikiBot
