Sipeed MaixStation Claims to Beat Nvidia DGX Spark at Just 40W

Dense performance versus sparse marketing

Nvidia loves to plaster headline TOPS numbers across marketing slides, advertising 1,000 sparse FP4 TOPS for its DGX Spark. Sipeed argues that real-world language model inference rarely translates to those synthetic theoretical peaks. Instead, the MaixStation relies on Axera's 6nm AX8860P processor, which delivers 256 dense TOPS alongside 16 Arm Cortex-A55 general-purpose cores.

According to Sipeed's internal benchmarks, the 128GB MaixStation slightly outperformed the DGX Spark on Qwen model evaluations. Both machines share a matched 273 GB/s memory bandwidth, but Sipeed insists its hardware sustains higher compute efficiency during the prefill stage of generation. The comparisons ran with speculative decoding and multi-token prediction disabled across both setups to ensure a level baseline.

Independent reviewers will naturally want to verify those claims across diverse model architectures. Still, seeing a 40W palm-sized enclosure match a machine consuming significantly more power highlights the growing pressure on mainstream accelerators, much like how regional buyers faced volatile pricing when RTX 5080 prices jumped up to $180 ahead of launch in China.

Sipeed MaixStation: specifications and hardware details

Specs, clustering, and software support

Squeezing a full-fledged local LLM rig into a 110 x 110 x 30 mm chassis required some clever engineering. The system runs off a standard 65W USB Power Delivery charger, features a 10GbE SFP+ port for fast networking, and incorporates Axera's proprietary AXLink interconnect.

That interconnect is where things get interesting for enthusiasts running massive open-weights models. Axera claims around 80% communication efficiency across clustered chips. A single 32GB model handles Qwen3.8-27B, while the 128GB top trim fits Qwen3.8-Flash-Next 125B. Daisy-chaining two units yields 256GB of memory to run DeepSeek V4-Flash 284B, and a four-box cluster reaches 512GB for DeepSeek V4.1-Flash 552B.

On the software side, Sipeed states out-of-the-box support for PyTorch, ONNX, vLLM, Hugging Face pipelines, and SGLang, alongside an OpenAI-compatible API endpoint for easy integration into existing developer tooling.

Availability and pricing

Sipeed plans to launch the MaixStation via a Kickstarter crowdfunding campaign in the coming months. Early buyers can secure a $100 discount through a $30 advance reservation.

The base 32GB model is expected to retail below $2,000, while the high-end 128GB model should land under $4,000. By contrast, Nvidia's DGX Spark has climbed well beyond its initial launch targets, reportedly commanding around $6,949 at retail.

If Sipeed delivers on its performance metrics and software maturity, the MaixStation could become one of the most compelling compact inference appliances on the market. We will be watching the upcoming Kickstarter launch closely to see if real-world hardware matches the bold claims.