Mac Studio M5 Ultra, Strix Halo or DGX Spark: which box really makes local language models fast

Apple announced the new Mac Studio with M5 Max and M5 Ultra on 25 August, shipping from 22 September, with the 512 GB variant at the end of October. That leaves three desktop boxes with 128 GB or more of unified memory sitting side by side, all promising the same thing: a large language model running locally, without a rack of graphics cards. We run exactly that on four RTX 5090s with tensor parallelism, so the question of whether one of these boxes could replace the setup is not a theoretical one. The answer sits in a single column of the spec sheet, and it is not the one with the petaflops.

The contenders

The prices are the first result. NVIDIA raised the DGX Spark list price on 23 February from 3,999 to 4,699 US dollars, citing the memory shortage; in Europe the price went from 4,180 to 4,800 euros, and the street price today is 5,629. The Framework Desktop with 128 GB still cost 1,999 US dollars at launch in February 2025, 2,459 in January 2026, and in the EU configurator it stands at 3,889 euros today, sold out. Company head Nirav Patel, speaking to PCWorld: it is precisely the 128-gigabit dies whose prices rose the most, and the 128 GB configuration takes eight of them. On the old Mac Studio, Apple temporarily dropped the 128 and 512 GB options for supply reasons.

Why bandwidth decides

Generating text pulls every active weight through memory once per token. A dense 70B model at 4-bit quantization is roughly 40 GB per token; at 273 GB/s that is seven tokens per second in theory, and less in practice. An MoE model with 3 billion active parameters pulls only two gigabytes per token. The measurements on the DGX Spark show exactly that: a 32B dense model delivers 10.7 tokens/s under llama.cpp, a 30B MoE with 3 billion active parameters 89.3, at the same quantization on the same machine. The Strix Halo shows the same picture, only sharper: a dense 70B model reaches 5.0 tokens/s, a 122B MoE with 10 billion active parameters 19.1.

The M5 Max figures come from the MacBook Pro with the same chip and the same 614 GB/s of bandwidth; hardware-corner measured them under MLX in March. For the M5 Ultra there is not a single measurement, because the machine has not shipped. Anyone deriving a number from the 1,200 GB/s is calculating, not measuring.

Where Apple loses

Bandwidth is not everything, and Tom's Hardware demonstrated that in July in a three-way comparison using the M4 Max. The Mac had twice the bandwidth of GB10 and Strix Halo, but delivered only 25 to 126 percent more during decode, depending on the model. During prefill, meaning reading in a long prompt, it fell behind the GB10 and on a dense 12B model even behind the Strix Halo. The older extreme case: an M3 Ultra needed 888 seconds to first token for an 8,000-token prompt against DeepSeek V3. Apple's answer in the M5 is matrix units in every GPU core, with a promise of 4× prompt processing over the M3 Ultra; MLX uses them, while the llama.cpp runtime in LM Studio does not yet, according to a bug report open since June.

Then there is the ecosystem. On the Mac there is no vLLM, no CUDA and no GPU inside a container. MLX is fast, but it is a path of its own. The DGX Spark is the only one of the three boxes where vLLM, SGLang, TensorRT-LLM and fine-tuning run without detours; the friction sits in ARM64 wheels, not in CUDA. Strix Halo is the only candidate with Windows 11, runs well with llama.cpp, but has no native FP8 in its RDNA 3.5 GPU, which is why vLLM exists there only as a self-built affair.

Where NVIDIA does not deliver what the spec sheet says

The DGX Spark petaflop is, per NVIDIA's own footnote, a theoretical FP4 figure with sparsity. In October, John Carmack measured the box levelling off at around 100 watts of power draw, less than half the 240-watt rating, and delivering roughly half the advertised performance; Tom's Hardware put the finding and NVIDIA's footnote side by side. At idle the Spark draws 35 watts, three to four times what a Strix Halo machine needs at 8 to 10 watts, mostly because of the ConnectX-7 network card. That card is required for the thing that sets the Spark apart from the others: two units can officially be linked.

Apple's equivalent, since macOS 26.2, is RDMA over Thunderbolt 5 plus a library for distributed MLX; at WWDC a trillion-parameter model ran on four M3 Ultras, independently re-measured at 28 to 30 tokens/s for Kimi K2 on four Mac Studios with 512 GB each.

What a gigabyte costs

The most striking direct comparison sits in the middle of the table: an HP Z2 Mini G1a with Strix Halo at 5,354 euros against a Mac Studio M5 Max with 128 GB at 5,859 euros. Nine percent more money for 2.4 times the bandwidth. The price advantage AMD held in spring 2025 has largely disappeared in this year's memory market; what remains of it sits with the Chinese mini PCs, which come with no support.

The successor at IFA

AMD has announced the Ryzen AI Max+ PRO 495 as the successor to the 395: the same 16 Zen 5 cores and 40 compute units, boost to 5.2 GHz, a 55 TOPS NPU, and above all up to 192 GB of LPDDR5X-8533, 160 GB of which the GPU can use. Bandwidth rises to 273 GB/s, exactly the DGX Spark level, 6.6 percent above the 395. Framework has announced a desktop built on it, and Acemagic is showing the F9A at IFA in Berlin, claiming it runs DeepSeek V4 Flash with 284 billion parameters locally. Nobody is naming prices. The arithmetic from above applies here too: 50 percent more capacity at six percent more bandwidth means you can load a larger model in order to run it more slowly. For MoE models with few active parameters that is still a real step forward; for dense models it is none.

Our view

For dense models from 30 billion parameters upwards, only the Mac Studio among the three is fast enough, and then only where MLX supports the model. For MoE models up to 120 billion parameters all three will do, and the ecosystem then decides: CUDA on the Spark, Windows and price on the Strix Halo, bandwidth and clustering on the Mac. None of them replaces our four RTX 5090s: 128 GB of VRAM at roughly 1.8 TB/s per card and 32 concurrent requests is a different class, at a different price, with different power draw. Anyone shopping for a single-workstation box should wait for the first M5 Ultra measurements after 22 September and for a price on the 495. Neither exists today, and both will change the ranking.

Further reading

• Apple Newsroom: Mac Studio with M5 Max and M5 Ultra
• Apple: Mac Studio tech specs
• NVIDIA: DGX Spark
• ComputerBase: AMD Ryzen AI Max+ 395 review
• Tom's Hardware: M4 Max against GB10 and Strix Halo
• LMSYS: DGX Spark with SGLang and Ollama
• llama.cpp discussion on the DGX Spark
• lhl: Strix Halo bandwidth and backends
• soothill.io: ROCm against Vulkan on Strix Halo
• Our post on DRAM bandwidth across 4× RTX 5090