Tom’s Hardware’s November 10, 2026 comparison found different leaders across its local-AI tests: a Mac Studio with M5 Ultra led one Qwen workload, while a two-system Spark setup was faster in the tested DeepSeek contexts. For GLM 5.3 Flash, changing the Mac’s inference engine also changed its result. The findings point to a workload-specific choice, not one overall winner. Tom’s Hardware’s comparison
Match the system to the model and inference engine you plan to run; a result from one setup doesn’t automatically carry over to another.
| Model and workload | Mac Studio with M5 Ultra | Spark or GB10 setup |
| Qwen 3.8 27B, AgentPerf-Local | llama.cpp with native multi-token prediction; finished in just under half the time of the GB10 and used about 16% less task energy. | One Dell Pro Precision GB10, described as nearly identical to DGX Spark; llama.cpp with native multi-token prediction. |
| DeepSeek V4.1 Flash 1017, longest tested contexts | 256GB system running llama.cpp; reported processing time of 10–25 minutes. The model weights and DSpark speculative-decoding model did not fit together, even at two-bit quantization. | Two-Spark recipe using TensorFold and DSpark speculative decoding; reported processing time under five minutes. |
| GLM 5.3 Flash | TensorFold with MLX-specific four-bit quantization averaged three times the decode throughput of the Mac’s llama.cpp result. | Two-Spark Mia’s AI Lab recipe; its prompt processing was faster than the Mac’s llama.cpp result. |
The Qwen result comes from a prompt-heavy test
For AgentPerf-Local, Tom’s Hardware ran dense Qwen 3.8 27B with llama.cpp and native multi-token prediction on the Mac and a single Dell Pro Precision GB10. The test replays recorded inference traces, and the outlet says most of the work is prompt processing: taking in the input context before generating a response. That makes the Mac’s result useful for this particular workload, but not a stand-in for every agent’s full sequence of tasks.
The single-system comparison used a Dell Pro Precision GB10, not an NVIDIA DGX Spark unit. Tom’s Hardware described the Dell as nearly identical to DGX Spark. That distinction matters when applying the result: the benchmark compared one Mac with one GB10 in this test, while the larger-model results used two Sparks.
DeepSeek favors a different system and recipe
In the longest tested DeepSeek V4.1 Flash 1017 contexts, Tom’s Hardware reported 10–25 minutes on the 256GB Mac and under five minutes on the two-Spark setup. The systems used different inference recipes, and the Spark result came from two machines. The Mac also could not fit both the model weights and the DSpark speculative-decoding model in memory, even at two-bit quantization.
That memory constraint and the software setup help explain why the result does not reduce to a simple hardware ranking. TensorFold supported the tested GLM setup but, at the time of Tom’s Hardware’s report, did not support the tested DeepSeek model. The outlet did not run end-to-end agentic tests for these larger models, so their reported inference timings are not measurements of complete agent workflows.
TensorFold changes the Mac’s GLM result
For GLM 5.3 Flash, the Mac’s TensorFold setup used an MLX-specific four-bit quantization. Tom’s Hardware reported average decode throughput three times that of the Mac’s llama.cpp result, with prompt processing closer to the Spark cluster. Decode throughput refers to generating output tokens; it is a different part of inference from processing the prompt. The charted TensorFold test used 2,048 input tokens and 512 output tokens.
The practical lesson is to check support for the specific model and engine together. A faster result with TensorFold on GLM does not predict the Mac’s result on DeepSeek, which the report said TensorFold did not support at that time.
What to weigh for several agents
Some users in a discussion about running multiple agents favor Spark for concurrent clients, citing CUDA batching with vLLM or SGLang. Another commenter reported splitting inference across Macs using Thunderbolt 5 and RDMA. These are community perspectives rather than controlled comparisons, so they can inform what to investigate in a multi-client setup but don’t establish which platform will handle a particular workload better. The discussion on Reddit
If you’re also choosing a Mac desktop for work beyond local AI, NeoTeo’s earlier Mac mini versus Mac Studio comparison covers that broader desktop decision.