A new 16-core AI chip from Axera is turning heads with desktop-class performance, generous memory, and a much lower projected price. The AX8860P could power compact systems from Sipeed and Radxa, giving developers a practical way to run large language models locally without stepping all the way up to Nvidia's increasingly expensive DGX Spark.
A small chip with surprisingly big ambitions
Axera's AX8860P is aiming squarely at the growing market for local AI workstations. The 6nm processor packs 16 Arm Cortex-A55 cores, a neural processing unit rated at up to 128 TOPS in standard configurations and 256 TOPS with FP4, plus support for up to 128GB of LPDDR5X memory.
That combination matters because running large language models locally is often less about raw peak compute than memory capacity and bandwidth. According to CNX Software, the chip can deliver up to 273GB/s, giving it a useful foundation for handling models that would otherwise need a much larger accelerator.
The package is also notably compact at 78 by 78 millimetres. With a typical load of around 40 watts, it looks less like a server component and more like the heart of a quiet desktop box. That's an appealing shift for developers who'd rather keep an AI system beside their monitor than in a rack.
Can it really challenge Nvidia's DGX Spark?
The comparison with Nvidia's DGX Spark is where the story gets interesting. Nvidia promotes Spark with a much higher headline figure of 1,000 TOPS, but that figure uses sparse FP4 performance. Axera argues that dense performance provides a fairer comparison for many real workloads, where theoretical peak numbers don't always translate neatly into useful output.
Sipeed's early comparisons reportedly place its 128GB MaixStation close to the 128GB DGX Spark in selected Qwen tests. The company says additional software tuning could improve the gap, although these figures should be treated as an early snapshot rather than a final verdict.
Both platforms are listed with 273GB/s of memory bandwidth, but Axera says its architecture is designed to keep more of that bandwidth active. That distinction sounds technical, but it can affect the experience users actually notice: faster responses, steadier token generation, and fewer moments where the system feels like it's waiting around.
MaixStation brings local AI to the desktop
Sipeed's MaixStation is expected to be one of the first practical products built around the AX8860P. The compact computer measures 110 by 110 by 30 millimetres, uses a 65W USB-C power supply, and includes networking and USB-C connectivity for use with a laptop or local network.
A 32GB version is expected to handle Qwen3.8-27B, while the 128GB model is intended for much larger models such as Qwen3.8-Flash-Next 125B. Linking two systems could provide 256GB of memory for a model such as DeepSeek V4-Flash 284B, while a four-node arrangement is aimed at models above 500B parameters.
Those claims will appeal to enthusiasts who want to experiment with local inference, but buyers should remember that model support isn't the same as effortless performance. Framework compatibility, quantisation, drivers, and thermal behaviour will all influence whether the finished product feels polished or merely promising.
Price could be the real advantage
Nvidia's DGX Spark launched with a price in the $3,000 to $4,000 range, but reported retail pricing has since climbed to nearly $7,000. Sipeed's expected pricing gives the MaixStation a considerably sharper angle, with the 32GB version projected below $2,000 and the 128GB version below $4,000.
That difference becomes even more important when systems are connected together. Sipeed says MaixStation units can be daisy-chained with a cable, whereas Nvidia configurations may require a costly high-speed network switch. For hobbyists, researchers, and small teams, the accessory bill can be just as important as the headline price.
The catch is timing. MaixStation isn't broadly available yet, and Sipeed plans to launch it through Kickstarter in the coming months. A reservation programme is already being discussed, but early buyers should wait for final specifications, shipping details, and independent testing before treating the projected price as a guaranteed bargain.
What developers should watch next
The AX8860P is designed to work with vLLM, SGLang, PyTorch, ONNX, Hugging Face formats, and OpenAI-compatible APIs. Axera has also shown an example workflow that converts a MobileNetV2 ONNX model into its own axmodel format, suggesting the company is building a complete software path rather than selling silicon alone.
Radxa is preparing a Pocket Supercomputer based on the chip, while Sipeed is working on PCIe cards that can combine as many as four AX8860P processors. Those products could broaden the platform beyond one compact computer and make it easier for developers to try different configurations.
For now, the sensible view is cautious optimism. The hardware looks unusually well matched to local AI, the power draw is modest, and the projected pricing is compelling. But the final test won't be a TOPS figure on a slide. It'll be how quickly a real model runs, how easily developers can deploy it, and whether the software stays out of the way.
Disclaimer: This article may have been created with AI assistance and reviewed by our editorial team. It is provided for general informational purposes only. Readers should verify information independently before relying on this content.
