Mac Studio M5 Ultra vs DGX Spark: 1.84x on 27B, Two GB10s Win DeepSeek

A single 256 GB Apple Mac Studio with M5 Ultra finished the AgentPerf-Local agent replay on Qwen 3.8 27B in 555.2 seconds, against 1,020.2 seconds for one Dell Pro Precision GB10, Tom's Hardware reported on Nov. 10.

That is the dense-model answer, and it is clear: on the same engine, the Mac did the same work 1.84 times as fast. On the large mixture-of-experts models that a 256 GB machine is bought for, the answer flips. A pair of GB10 boxes, the platform NVIDIA sells as the DGX Spark, took the lead on DeepSeek V4.1 Flash. But the two platforms ran different inference engines and different quantizations in those tests, so the gap measures the software stack as much as the hardware. This piece follows our Mac Studio M5 Ultra review of Nov. 1, which covered the first round of measurements.

Mac Studio M5 Ultra vs GB10 Specs and Bandwidth

Apple's spec page lists 1.2 TB/s of memory bandwidth for the 36-core CPU, 80-core GPU M5 Ultra, with 256 GB or 512 GB of unified memory. NVIDIA lists the DGX Spark at 128 GB of LPDDR5x on a 256-bit bus at 273 GB/s, with a ConnectX-7 port at 200 Gbps that links up to four units into one larger memory pool. Decode speed, the rate at which a model writes tokens, is mostly a function of how fast weights stream out of memory. So the Mac starts with just over four times the GB10's bandwidth, as Tom's Hardware notes. The quantized DeepSeek and GLM builds in this test fit in the Mac's 256 GB, but Tom's Hardware says they will not fit on a single 128 GB GB10.

The tested Dell box is the GB10 system Dell launched as the Dell Pro Max with GB10 and now lists as the Dell Pro Precision with GB10 (model FCM1253). Tom's Hardware calls it nearly identical to the DGX Spark, and our Dell GB10 review and our DGX Spark review cover the platform's earlier measurements.

Dense 27B: one Mac against one GB10

This is the only like-for-like test in the article. AgentPerf-Local comes from Artificial Analysis. Tom's Hardware ran llama.cpp on both machines and enabled Qwen 3.8 27B's native multi-token prediction (MTP), which drafts several tokens per step, on both. The benchmark plays back recorded agent traces and does a fixed amount of work, most of it prompt processing rather than token generation.

The Mac finished in 555.2 s and the GB10 in 1,020.2 s. Tom's Hardware also logged wall power across the run: 138.96 kJ for the Mac and 165.96 kJ for the Dell, about 16% less energy for the Mac. Dividing energy by time gives an average draw of about 250 W for the Mac and about 163 W for the GB10, by our arithmetic. The Mac pulls more power, but for so much less time that the job costs less energy. Its energy-delay product, time multiplied by energy, was 77,178.4 against 169,312.4, under half the GB10's.

Tom's Hardware added a caveat of its own: a different engine might favour the GB10. The result holds for llama.cpp on this model.

DeepSeek V4.1 Flash: two GB10s against one Mac

Here the engines differ, and the numbers below should be read with that in mind. The Mac ran Unsloth's 2-bit Q2_K_XL build on llama.cpp, the only engine Tom's Hardware could use for DeepSeek on Apple Silicon. It also had to run without speculative decoding, because 256 GB could not hold both the weights and the DSpark draft model. The two-GB10 cluster ran a community recipe from Mia's AI Lab: an EXL3 quantization on the TensorFold engine with DSpark speculative decoding.

With 2,048 input and 512 output tokens on top of the context, time to first token on the Mac rose from 3.9 s at zero context to 273.9 s at 131,072 tokens, 614.8 s at 262,144 and 1,509.1 s at 524,288. The cluster's charts show 1.2 s, 65.9 s, 133.7 s and 311.3 s at the same points. Against the Mac's times that is 31% at zero context and 21% to 24% from 131,072 tokens up. Decode favoured the cluster too: 66.2 tok/s at zero context and 62.7 at 524,288, against 42.9 and 34.0 on the Mac. Tom's Hardware's verdict is that a pair of Sparks is the clear winner for this model today.

GLM 5.3 Flash: the engine decides

GLM 5.3 Flash is the test that shows how much of the DeepSeek gap is software. On llama.cpp with Unsloth's UD-Q4_K_XL build and MTP, the Mac took 60.5 s to first token at 32,768 tokens and 484.6 s at 262,144. The cluster, on TensorFold with an EXL3 4-bit build, took 19.5 s and 154.4 s. On decode the Mac led at short context, 46.3 tok/s against 36.5 at zero and 43.7 against 37.8 at 32,768, but by 131,072 tokens the cluster was level (38.2 against 37.2).

Then Tom's Hardware ran the Mac on TensorFold with an MLX 4-bit build and MTP. Time to first token fell to 16.9 s at 32,768, 137.1 s at 262,144 and 295.4 s at 524,288. Decode rose to 94.0 tok/s at zero context and held 79.4 at 524,288, which Tom's Hardware calls "three times faster, on average" than llama.cpp; the charts' shared points give about 2.0 to 2.8 times. On the charts, those Mac prefill times sit slightly below the cluster's at every context depth, and decode runs 1.4 to 3.4 times the cluster's, depending on context depth. Both now use TensorFold, but the quantization formats differ (MLX against EXL3), and the Mac had MTP on, so this is still not a strict match. Tom's Hardware describes the Mac's prefill as getting much closer to the cluster's.

TensorFold does not support DeepSeek on Apple Silicon yet. Until it does, the Mac's best result on one of the two big models comes from a slower engine.

Price: closer than it looks

Tom's Hardware priced its tested Mac Studio at $12,300 or more. Dell.com listed the Dell Pro Precision with GB10 at $7,901.00 on Nov. 10, with 128 GB and a 2TB QLC drive. The 4TB TLC drive in the tested unit adds $2,450.00, for $10,351, in line with the roughly $10,000 Tom's Hardware cites. Two GB10 boxes with 4TB drives run $15,000 to more than $20,000, the outlet says, depending on the maker. At the low end that is 25% more than the roughly $12,000 Mac, Tom's Hardware says (about 22% against the $12,300 minimum), and our earlier GB10 price comparison shows how fast those prices have moved. Lenovo set all ThinkStation PGX systems to "coming soon" during the testing, according to Tom's Hardware.

The Mac's catch is time. Tom's Hardware puts lead times at 13 to 15 weeks, into February 2027. The 512 GB model starts at $22,299, as MacRumors reported, and none has been measured yet.

Why Bandwidth Alone Does Not Set Local AI Speed

A model's answer has two phases. Prefill reads the prompt and is limited mostly by compute. Decode then writes one token at a time, pulling the model's weights from memory for each step, so memory bandwidth limits it. Apple lists 1.2 TB/s for the M5 Ultra and NVIDIA lists 273 GB/s for the DGX Spark, which is why the Mac leads decode on GLM 5.3 Flash.

Software can still move the result more than the hardware gap. Tom's Hardware measured the Mac at 46.3 tok/s on llama.cpp and 94.0 tok/s on TensorFold at zero context, with a different quantization and multi-token prediction on the second run. For a buyer, the model and the engine recipe decide speed as much as the box does.

Which to buy

For a single developer running a dense model of the Qwen 3.8 27B class, the Mac Studio is the measured winner: 1.84 times as fast as one GB10 on the agent replay, with about 16% less energy, on the same engine.

For a buyer who wants DeepSeek V4.1 Flash locally now, two GB10 boxes are the measured choice, with a fifth to a quarter of the Mac's wait at long context and faster decode. The cost is more money, a second box and a cluster to set up. The cluster's lead also rests on a community recipe and on TensorFold support that the Mac lacks.

For GLM 5.3 Flash, the Mac on TensorFold matched or beat the pair on Tom's Hardware's charts, at a lower price. That holds only for a model with an Apple Silicon recipe. Tom's Hardware's view is that the NVIDIA platform's broader engine support means a new model usually gets a working recipe sooner, and NVIDIA's own forums show the GB10 still has open kernel issues, as we reported on Nov. 9.

Neither machine fits a buyer who needs a large MoE model on one box today without waiting. The Mac's lead time runs into 2027, and a single 128 GB GB10 cannot hold these models.

What is not confirmed yet

No 512 GB M5 Ultra has been independently measured; first orders arrive Dec. 3, according to Digital Trends. Tom's Hardware reported no wall power for the cluster in the DeepSeek and GLM tests, so energy per task is known only for the 27B run. Its text says the cluster needs under five minutes for the longest DeepSeek contexts, while its chart shows 311.3 s at 524,288 tokens. The cluster's DeepSeek decode chart is labelled llama.cpp, while the text says TensorFold. Tom's Hardware has not explained the cluster's GLM decode jump to 63.2 tok/s at 262,144 tokens. The next fact to watch is TensorFold support for DeepSeek on Apple Silicon, for which no date has been given.