Lenovo RTX Spark Overtakes Mac Studio in Local AI Coding

Summary

ZDNET writer David Gewirtz pitted the €3,720 Lenovo Yoga 9n (Nvidia RTX Spark NX1, 64 GB RAM, Blackwell GPU with 6,144 CUDA cores) against a more expensive €5,500 Mac Studio M4 Max in local AI coding tests. In the test, the laptop was approximately 3 times faster in prompt processing; however, when the 64 GB memory limit was reached, it remained 1.5 to 2.5 times slower in token generation.

Why it matters

This is one of the first real-world, side-by-side tests of Nvidia’s new SoC-class GPU for on-device LLM inference. Apple had held a quiet monopoly in this area. The results show that a portable device under €4,000 can genuinely compete with a desktop workstation in AI coding. This could change developers’ thinking about local inference hardware. On the other hand, the hard 64 GB memory limit and 32,000-token context window also reveal just how far local models are from cloud-scale workflows.

Highlights

• Lenovo was approximately 3 times faster than the Mac Studio in prompt processing. However, it lagged by 1.5 times on the 8B model and 2.5 times on the 32B model in token generation. The author attributes this gap to memory saturation.
• The local coding stack (OpenCode + Qwen3-Coder-Next 80B + llama.cpp/CUDA) was limited to a 32,768-token context window. This is about one-thirtieth of the capacity routinely used by cloud agents like Claude Code.
• Under sustained load, the Yoga’s fan becomes noticeably loud. The 64 GB RAM limit cannot be upgraded. This creates a decisive constraint for larger models.

Critical take

The article’s claim in the title that the Spark 'outperformed' the Mac Studio is only half true. The laptop leads in prompt ingestion; however, it loses in the token generation phase where code is actually produced. The author’s own 32,000-token context limit shows that the 'coding test' is only a small part of a real agentic session. Because the review cuts off midway, the reader sees neither a final verdict nor a cost-per-token comparison. This leaves the 'we were impressed' enthusiasm hanging in the air.