Sipeed MaixStation Pitches a $3,999 Local AI Box With 128GB LPDDR5X

Sipeed’s MaixStation is one of the more interesting local-AI systems to surface in October 2026 because it aims for the same memory class as NVIDIA’s DGX Spark, but in a much smaller and potentially cheaper package. As of Sunday, October 18, 2026, this is not a shipping product: Sipeed lists a Kickstarter launch for the end of 2026, while Tom’s Hardware reports that units are expected after that campaign closes. That matters, because the MaixStation is best viewed as a promising workstation concept rather than a proven DGX Spark replacement. (sipeed.com)

The headline configuration is built around Axera’s AX8860P, with Sipeed listing 256 TOPS FP4, 128GB LPDDR5X, a 256-bit LPDDR5X-8533 memory interface, and 273 GB/s of memory bandwidth. Sipeed’s own page also describes the Maix-V compute module as a 78 × 78 mm board using an LPCAMM2 connector, shared across MaixStation, MaixStation Pro, and MaixBlade designs. Tom’s Hardware further reports a 16-core Arm Cortex-A55 CPU, a 6nm process, and a typical box size of about 11 × 11 × 3 cm with around 40W typical draw. (sipeed.com)

Compare with ASUS GeForce RTX 5090 for a GPU-based local AI build

On memory alone, the comparison is easy to understand: the MaixStation’s 128GB LPDDR5X and 273 GB/s bandwidth match the key memory figures NVIDIA lists for the current DGX Spark, which also has 128GB of coherent unified memory. The price story is where the pitch gets sharper. NVIDIA’s marketplace currently lists the DGX Spark at $6,950, while Sipeed told Tom’s Hardware it expects the 128GB MaixStation to come in around $3,999, with Kickstarter launch discounts possible. That makes the MaixStation look like a lower-cost alternative on paper, but not a like-for-like substitute yet. NVIDIA ships a mature CUDA-based software stack, while Sipeed’s hardware and NPU software still need wider validation. (marketplace.nvidia.com)

See NVIDIA DGX Spark Personal AI Desktop Supercomputer price

Sipeed is positioning MaixStation for local inference rather than general workstation graphics. The company says the box will ship with Debian 13 and Linux 6.18, and claims support for vLLM, SGLang, PyTorch, llama.cpp, and Ollama through NPU acceleration or custom backends. Sipeed’s FAQ says SGLang and vLLM use a custom hardware-acceleration backend with kernels written in AXDSL, and Tom’s Hardware reports that the broader NPU toolchain is due by the end of October 2026. Until developers can test model compatibility, operator support, and real throughput, these software claims should be treated as part of the pitch rather than confirmed ecosystem maturity. (sipeed.com)

View Lenovo Legion Pro 7i laptop with Core Ultra 9 and NVIDIA RTX 5090 24GB on partner website

The MaixStation is also notable for how Sipeed wants users to connect and scale it. The front panel shown on Sipeed’s page includes a power button, power USB-C, USB 3.0, a mode switch, and 10G SFP+. Tom’s Hardware additionally reports an M.2 2280 slot for a PCIe 5.0 SSD and an MCIO 8i connector for linking a second unit. Sipeed is also talking about larger setups through MaixStation Pro and MaixBlade, with 32GB, 128GB, and 512GB-class options appearing in its materials. For users interested in running local models without renting cloud GPUs, the concept is attractive: lots of memory, modest power, and local data control. The important caveat is that the most exciting parts — final pricing, delivery timing, real model performance, and software polish — are still waiting on the Kickstarter and independent testing. (sipeed.com)

See Crucial T705 13 GB/s NVMe SSD price