Sipeed lists model-specific Qwen throughput claims for the MaixStation, its announced local-AI system built around Axera AX8860P. The figures compare two FP4 workloads with NVIDIA DGX Spark, with MTP disabled. Tom’s Hardware reported the system’s reveal on October 13, 2026; Sipeed’s product page carries the detailed comparisons.
The Qwen figures Sipeed publishes
Sipeed’s comparison gives Prefill and Decode rates in tokens per second. Prefill measures processing the prompt; Decode measures generating output. The product page labels the workloads dense FP4 and MoE FP4, and says MTP is disabled. These are Sipeed’s published comparisons.
| Model and mode | Metric | NVIDIA DGX Spark (tokens/s) | Sipeed MaixStation (tokens/s) |
| Qwen3.8-27B, dense FP4 | Prefill | 1,300 | 1,400–1,600 |
| Qwen3.8-27B, dense FP4 | Decode | 13 | 15–17 |
| Qwen3.6-35B-A3B, MoE FP4 | Prefill | 5,600 | 6,300–7,700 |
| Qwen3.6-35B-A3B, MoE FP4 | Decode | 80 | 80–88 |
Hardware, software and campaign plans
For its headline configuration, Sipeed lists 128GB of LPDDR5X memory, 273 GB/s of memory bandwidth, a 256 TOPS FP4 rating and 40W typical load. The listed dimensions are 11 × 11 × 3 cm. Sipeed also lists 32GB and 128GB MaixStation configurations, plus 512GB-or-higher configurations associated with MaixStation Pro and MaixBlade.
Sipeed lists Debian 13 and Linux 6.18, and names vLLM, SGLang, PyTorch, llama.cpp and Ollama as supported inference frameworks. The company says it is preparing the AX8860 SDK for open sourcing.
Sipeed plans to launch a Kickstarter campaign for MaixStation at the end of 2026.
