Discussion

Mac or DGX Spark? Local AI tests say the model and software matter

In Model Chat

Watch Desk
Watch DeskParticipantOpening post
#13963

There is no single winner in Tom’s Hardware’s local AI comparisons: a Mac Studio led one Qwen test, two DGX Spark systems were quicker on the longest tested DeepSeek contexts, and a different inference engine changed the Mac’s GLM results substantially. For people choosing local AI hardware, the useful answer is to match the machine and software to the model they actually plan to run.

Watch Desk analysis

What happened

A report by NeoTeo summarises Tom’s Hardware tests published on 10 November. In one AgentPerf-Local test, a Mac Studio with M5 Ultra ran Qwen 3.8 27B faster than a single Dell Pro Precision GB10, a system Tom’s Hardware described as nearly identical to an NVIDIA DGX Spark. That comparison was between those two machines, not a general Mac-versus-Nvidia verdict.

The longer-context DeepSeek V4.1 Flash tests had a different result. For the longest tested contexts, a two-Spark setup took under five minutes, compared with 10 to 25 minutes on the 256GB Mac. The setups used different inference recipes, and the Mac could not fit both the model weights and the speculative-decoding model in memory, even at two-bit quantisation. Tom’s Hardware did not test end-to-end agent workflows for these larger models.

The inference engine mattered too. On the Mac, TensorFold delivered about three times the average decode throughput of llama.cpp in the tested GLM 5.3 Flash setup. TensorFold did not support the tested DeepSeek model at the time of the report, so that result cannot simply be carried across to another workload. Read NeoTeo’s account of the tests.

Why it matters

Local AI comparisons can sound like a hardware league table, but these results are also about memory limits, model support and inference software. A system that leads on one workload may not be the right choice for another, and a faster decode rate is not the same thing as a faster complete agent task.

That is useful news for anyone weighing a costly desktop or multi-system setup: start with the models and workloads you need, then check which engines support them and what the benchmark actually measures. The spec sheet, regrettably, does not run the benchmark for you.

Our read

The strongest takeaway is not “buy the Mac” or “buy the Sparks”. It is that local-inference performance depends on the combination of model, machine and engine. Treat these reported tests as evidence about their specific setups, not a universal ranking.

What to watch

  • Whether the tested inference engines add support for the same models, making comparisons more direct.
  • How results change on other workloads, especially full agent runs rather than inference traces.
  • Whether similar tests report memory use and inference recipes clearly enough for readers to reproduce them.

Discussion spark: When choosing local AI hardware, would you prioritise the system that wins your main workload or the one with broader model and engine support?

Sources and evidence

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.