Sipeed MaixStation: A $3,999 128GB Box for Local LLMs

Sipeed MaixStation is shaping up to be one of the most interesting new boxes for running local LLMs at home. Revealed this week, the compact system pairs Axera's new AX8860 AI chip with 128GB of LPDDR5X memory, and Tom's Hardware reports it is expected to cost $3,999. That puts a large-memory local AI machine into the same conversation as NVIDIA's DGX Spark and AMD's Ryzen AI Halo, with a very different chip inside.

• Chip: Axera AX8860 with 16 Arm Cortex-A55 cores and an NPU rated at 256 TOPS (FP4, no sparsity), built on a 6 nm process.
• Memory: 128GB of LPDDR5X on a 256-bit bus with up to 273 GB/s of bandwidth.
• Power and I/O: about 40 W for the AI module, PCIe 5.0 and a 256 Gbps networking interface, according to Wccftech.
• Price and timing: expected at $3,999, with units shipping after a year-end Kickstarter, per Tom's Hardware.

What Is the Axera AX8860?

The AX8860 is Axera's biggest AI processor yet. Its 16 Cortex-A55 cores handle general housekeeping, while the NPU does the heavy lifting for inference. Wccftech reports the 256 TOPS rating is measured in FP4 with no sparsity, which matters because many competing figures assume sparse math and look larger on paper.

The chip also scales out. Sipeed says two AX8860s reach about 512 TOPS and four reach about 1,024 TOPS, with roughly 80% scaling efficiency. Beyond the MaixStation desktop, Wccftech describes a single-slot PCIe card version and a quad-chip card, with smaller mini PCs reportedly in development.

How Much Memory Do Local LLMs Need?

For large language models, memory capacity and bandwidth matter as much as raw compute. A 128GB pool means the MaixStation can hold quantized models far larger than any consumer graphics card, and its 273 GB/s bandwidth matches the class of unified-memory machines like NVIDIA's DGX Spark. If you are still working out your own requirements, our guide on how much RAM a local LLM needs breaks down model size by quantization level.

Software support is also encouraging. Sipeed says the platform supports vLLM, SGLang and PyTorch, and Tom's Hardware reports the NPU toolchain is due at the end of this month.

How Does MaixStation Compare to DGX Spark?

Sipeed claims 25% faster prefill and 30% faster decode than DGX Spark when running a Qwen 27B model. Those are the company's own numbers and have not been independently tested yet, so treat them as an early signal rather than a verdict. What is clear from the spec sheet is the efficiency story: a roughly 40 W AI module with 128GB of memory is a compelling recipe for an always-on home AI server.

Sipeed is no stranger to this space, either. The company runs the llmdev.guide benchmark site that compares mini PCs for local AI, so it knows exactly which numbers buyers look at.

When Can You Buy a MaixStation?

Not quite yet. According to Tom's Hardware, Sipeed plans a Kickstarter campaign at the end of the year, with units shipping after it closes. Wccftech reports Sipeed is targeting availability within the year for devices built on the chip. More competition in 128GB local AI hardware is good news for everyone who wants to run serious models without the cloud. Keep an eye on our mini PC coverage for hands-on results once units land.

Sources: Tom's Hardware, Sipeed reveals DGX Spark and Ryzen AI Halo rival, expected to cost $3,999 — October 15, 2026; Wccftech, Axera AX8860 AI chip with 128 GB memory at 40W — October 13, 2026.