VarwoSipeed MaixStation Qwen Benchmarks: Claims and Comparison Limits
Understand MaixStation’s Qwen comparison settings, distinguish vendor claims from independent results and check what a reproducible test needs.
The MaixStation versus DGX Spark comparison is currently a Sipeed-reported claim, not a benchmark reproduced by Varwo. Begin with Sipeed's MaixStation page, checked October 31, 2026.
What the source specifies
The page identifies an AX8860P platform with 256 TOPS FP4, 128GB LPDDR5X and 273 GB/s bandwidth for the illustrated system. Its Qwen charts name Qwen3.8-27B dense and Qwen3.6-35B-A3B MoE, FP4 and MTP disabled. Prefill and decode are separate measures.
Both our text reader and browser observation displayed chart labels with 0/0–0 placeholders. These are not valid observed throughput results. We cannot substantiate an exact percentage advantage from those rendered values, and do not turn a promotional comparison into an independently measured score.
The page describes a planned end-of-2026 Kickstarter and indicative pricing rather than a verified delivered retail configuration. We did not reserve a unit, pay a deposit or confirm final availability.
Compare like with like
A meaningful Qwen test needs the same checkpoint, quantization, context and output lengths, batch/concurrency settings and decoding options. Record runtime versions and whether the numbers are a measured build or a tuning target.
| Measure | Question to ask |
|---|---|
| Prefill | How long is the input and what batch is used? |
| Decode | Is it single-stream, and is speculative decoding enabled? |
| Power | Which load, meter and system boundary are used? |
| Practical output | Does the model complete your task correctly? |
TOPS alone cannot answer those questions. Varwo has not run either device or independently verified performance. For small-model evaluation planning, see Liquid LFM3.