Where the numbers come from. Mac figures come from the LLMCheck index, the same per-Mac figures as its Best by Mac pages, each with its provenance mark: a filled square is a published figure, a hollow circle a community run, a diamond an estimate. The DGX Spark and Strix Halo figures are not in the index, which covers Apple Silicon only. They are quoted from the sources linked beside them, with the runtime and date. LLMCheck runs no tests of its own. The machines ran different runtimes and builds, so treat gaps of 10% or less as ties.
The short answer
- Fastest text generation at 128 GB: the Mac. A Mac Studio or MacBook Pro with M5 Max reads memory at 614 GB/s. The DGX Spark manages 273 GB/s and a Strix Halo box about 256 GB/s. On every model in the decode table below, the M5 Max figure is the highest of the three.
- Fastest prompt processing of the two PCs: the DGX Spark. On gpt-oss-120b it reads a prompt three to four times as fast as Strix Halo on an empty context, and two to five times as fast at 32K tokens. Agent tools re-read long prompts all the time, and there the gap matters.
- Cheapest 128 GB machine: Strix Halo, though no longer cheap: the Framework Desktop rose from $2,459 in January 2026 to $3,449 by September, and was out of stock on 18 November. It generates text about as fast as the Spark. The cheapest machine here that holds gpt-oss-120b is a 64 GB Mac mini M5 Pro ($2,699).
- Anything above 128 GB in one box: only the Mac. A Mac Studio M5 Ultra comes with 256 or 512 GB. Two DGX Sparks cable together to make 256 GB.
Memory, bandwidth and price
| Mac Studio M5 Max | NVIDIA DGX Spark | AMD Strix Halo (Ryzen AI Max+ 395) | |
|---|---|---|---|
| Memory | 128 GB unified | 128 GB LPDDR5x, unified (a 64 GB version is sold through OEMs) | Up to 128 GB LPDDR5X-8000. The GPU gets 96 GB on Windows (AMD) and more on Linux |
| Memory bandwidth | 614 GB/s (Apple) | 273 GB/s (NVIDIA) | About 256 GB/s: 256-bit at 8,000 MT/s. AMD publishes no figure |
| GPU | 40-core Apple GPU | Blackwell with 5th-gen Tensor Cores; NVIDIA claims up to 1 PFLOP of FP4, with sparsity | Radeon 8060S, 40 compute units |
| Operating system | macOS | DGX OS (based on Ubuntu) | Windows or Linux |
| Price | $5,099 (Apple list price, checked 14 Nov 2026). 128 GB MacBook Pro: $6,699 | $3,999 at launch (Oct 2025), then $4,699 from Feb 2026 (NVIDIA). The Register reported $6,950 on 2 Oct 2026; NVIDIA’s own pages showed no price on 18 Nov | Framework Desktop: $2,459 (Jan 2026), $3,449 by Sep 2026, out of stock on 18 Nov. GMKtec EVO-X2: $3,649 with 2 TB, about $3,499 with 1 TB (Jul 2026). AMD Ryzen AI Halo: $3,999.99 (Jun 2026) |
Sources:
- NVIDIA’s DGX Spark specifications.
- NVIDIA’s price-change notice (25 Feb 2026) and The Register (2 Oct 2026).
- NVIDIA’s 64 GB announcement (2 Oct 2026): from $4,999 through OEMs, on sale from 23 Oct.
- AMD’s 96 GB VGM post (29 Jul 2025). AMD’s ROCm tuning guide raises the Linux GPU allocation with
amd-ttm; its example gives the GPU 100 GB of a 128 GB system. - Strix Halo prices: Tom’s Hardware (Framework, 13 Jan 2026), Notebookcheck and Hardware Busters (Framework 128 GB at $3,449, 30 Sep 2026), Notebookcheck (GMKtec, 18 Jul 2026) and TweakTown (Micro Center, 14 Jun 2026).
Dell’s GB10 workstation (128 GB, 2 TB) was $7,901 on Dell’s US store on 18 Nov 2026. Memory prices have pushed every one of these machines up during 2026, so check the price on the day you buy.
Decode: the Mac’s lead
Decode is the speed at which a model writes its answer, in tokens per second. Each new token reads the model’s active weights once, so bandwidth sets the ceiling. This table shows plain decode at short context, with no speculative decoding:
| Model | Mac Studio M5 Max | Mac Studio M5 Ultra | DGX Spark (not in the index) | Strix Halo (not in the index) |
|---|---|---|---|---|
| gpt-oss-120b (MXFP4) | community92 tok/s | sourced130 tok/s | 58.7 tok/s (llama.cpp b7941, CUDA) | 56.6 tok/s (llama.cpp, Vulkan RADV); 51.8 on ROCm 7.2.3 |
| gpt-oss-20b (MXFP4) | community141 tok/s | community184 tok/s | 83.4 tok/s (llama.cpp b7941) | 79.8 tok/s (Vulkan RADV) |
| Qwen 3.6-35B-A3B (4-bit) | community130 tok/s | community153 tok/s | No accountable figure found | 60.4 tok/s (UD-Q4_K_XL, Vulkan RADV) |
| Llama 70B class (4-bit) | sourced13 tok/s | sourced25 tok/s | 4.42 tok/s (Llama 3.1 70B q4_K_M, Ollama 0.12.6) | About 5.0 tok/s (Shisa V2 70B, a Llama 3 70B fine-tune, i1-Q4_K_M, Vulkan, community) |
Sources for the non-Mac figures:
- DGX Spark: Georgi Gerganov’s llama.cpp benchmark file (tg32, build b7941, run 5 Feb 2026) and the Ollama blog (23 Oct 2025).
- Strix Halo: Donato Capitella’s Strix Halo toolboxes (tg128 on a Framework Desktop 128 GB, llama.cpp b9187, b9193 for the Qwen run, data from 16 May 2026). The 70B figure is one Level1Techs forum user’s run of Shisa V2 70B, a Llama 3 70B fine-tune, at i1-Q4_K_M on a pre-production Framework Desktop (22 Jul 2025).
- Mac Llama row: the index’s figure is Llama 3.3 70B.
The pattern holds wherever all three can run a model. On gpt-oss the DGX Spark and Strix Halo land within about 5% of each other, matching their near-identical bandwidth. An M5 Max Mac with twice that bandwidth lands well above both. NVIDIA’s own blog (24 Oct 2025) puts the Spark at 55.4 tok/s on gpt-oss-120b with llama.cpp, close to Gerganov’s figure. A cheaper Mac also gets into the same range: the 64 GB Mac mini M5 Pro ($2,699) holds gpt-oss-120b with about 2 GB to spare, and the index puts it at estimated~54 tok/s.
Speed falls as the context fills. At 32K tokens of context, Gerganov’s Spark run decodes gpt-oss-120b at 42.8 tok/s and Capitella’s Strix Halo run at 43.0 on Vulkan RADV (36.3 on ROCm). The index records Mac figures at the shortest context only, so it has no like-for-like Mac number yet.
Prompts and agents: the Spark’s lead
Prompt processing, or prefill, is how fast a machine reads the text you send before it starts answering. It is limited by GPU compute, not bandwidth, and here the ranking flips:
- Empty context. On gpt-oss-120b, Gerganov’s DGX Spark run reads 2,444 tok/s (pp2048). Capitella’s Strix Halo runs read 720 tok/s on Vulkan and 626 on ROCm. His test uses a shorter 512-token prompt (pp512).
- 32K tokens already in context. The Spark reads 1,567 tok/s. Strix Halo reads 308 on Vulkan and 653 on a ROCm 7 nightly build.
The LLMCheck index does not record prefill yet, so this page has no Mac prefill figure.
That matters most for coding agents, which send tens of thousands of tokens of context with every step. Artificial Analysis’s AA-AgentPerf-Local benchmark (29 Sep 2026) replays real agent sessions. It runs every machine with speculative decoding, so it is a different workload from the decode table above. On its laptops and workstations board, the DGX Spark finished fastest of the unified-memory machines; an RTX 5090 desktop on the same board was faster on every model that fits its 32 GB:
| Model (AA-AgentPerf-Local, minutes) | MacBook Pro M5 Pro, 64 GB | AMD Ryzen AI Halo, 128 GB | DGX Spark, 128 GB |
|---|---|---|---|
| Qwen3.8 27B | 37.6 | 34.5 | 17.3 |
| Qwen3.6-35B-A3B | 11.9 | 11.7 | 7.3 |
| Ling 3.0 Flash | does not fit | 25.1 | 14.9 |
The only Mac on that board is an M5 Pro, the 307 GB/s chip. In its M5 Ultra review (21 Sep 2026), Tom’s Hardware found that Mac’s prompt speed beat the Spark on Qwen 3.8-27B, with decode “almost four times higher”. In a follow-up (10 Nov 2026), a 256 GB, 80-core M5 Ultra Mac Studio finished the AgentPerf-Local Qwen 3.8-27B run in just under half the time of one Dell GB10. Two clustered GB10s won on prompt speed for the larger DeepSeek V4.1 Flash and GLM 5.3 Flash. Those results are published mostly as charts, so this page quotes no numbers from them.
Software
- DGX Spark: CUDA, with all its tools. NVIDIA’s Spark playbooks cover vLLM, SGLang, llama.cpp, NIM, LM Studio, NVFP4 quantization and multi-Spark setups. If your work has to run on NVIDIA servers later, this is the box that matches them.
- Strix Halo: improving, still rough. AMD’s ROCm compatibility matrix (10.2, 18 Nov 2026) lists the chip (gfx1151) as supported on Ubuntu and Windows, as it has since 10.0 (26 Aug 2026). Capitella calls llama.cpp on Vulkan the most stable option. Tom’s Hardware (6 Jul 2026) found the AMD Ryzen AI Halo’s preinstalled vLLM failed on Qwen 3.6-35B-A3B and fell back to llama.cpp.
- Mac: MLX, Apple’s own framework, plus llama.cpp on Metal, LM Studio, Ollama and oMLX. Ollama has used MLX by default on Apple Silicon since version 0.40.0 (6 Oct 2026), for the models its MLX engine supports. There is no CUDA, so CUDA-only research code will not run.
Models above 128 GB
All three top out at 128 GB in one box. The Mac goes further:
- Mac: a Mac Studio M5 Ultra comes with 256 GB ($9,499) or 512 GB ($22,299), Apple list prices checked 14 Nov 2026. That is enough for MiMo-V2.6-Flash and the other 192–256 GB picks. Macs can also be clustered: macOS 26.2 and later support RDMA over Thunderbolt 5 (Apple TN3205).
- DGX Spark: two units connect with one QSFP cable to make 256 GB for models of up to about 400B parameters. Four need a 200 GbE switch (NVIDIA).
- Strix Halo: stops at 128 GB. Its successor, the Ryzen AI Max+ PRO 495, takes up to 192 GB at about 273 GB/s (ServeTheHome, 20 May 2026), and its first systems were listed from about $6,800 (GMKtec and Framework, late Sep 2026; Minisforum’s is $7,399, Club386, 6 Oct 2026).
To compare two Macs at these sizes, see Mac Studio M5 Max 128 GB vs M5 Ultra 256 GB.
Which to buy
- You chat and code with 30–120B models and want the fastest replies: a Mac with M5 Max and 128 GB. It costs $5,099 as a Mac Studio or $6,699 as a MacBook Pro, and it is quiet. See its ranked model list.
- Your budget is under $3,000: a Mac mini M5 Pro with 64 GB ($2,699), the cheapest machine here that holds gpt-oss-120b, with little room left for context. A 128 GB Strix Halo box now costs more than $3,000 (the Framework Desktop $3,449 by September) and decodes gpt-oss-120b at about the same speed: more memory, but a rougher stack and slow prompt processing.
- You run coding agents with very long contexts, or you need CUDA: the DGX Spark. Check its current price first: NVIDIA raised it to $4,699 in February, and The Register reported $6,950 in October.
- You want more than 128 GB in one machine: a Mac Studio M5 Ultra, or two DGX Sparks cabled together.
Not sure which Mac? The Mac Advisor turns a budget and a use case into an exact configuration. The Mac-vs-Mac pages put any two configurations side by side.
Frequently Asked Questions
Is a Mac faster than an NVIDIA DGX Spark for local LLMs?
For generating text, an M5 Max or M5 Ultra Mac is. It reads memory at 614 or 1,228 GB/s against the Spark's 273 GB/s, and the LLMCheck index's M5 Max figure for gpt-oss-120b is well above the 58.7 tok/s published for the Spark with llama.cpp. For reading long prompts the Spark is far faster than Strix Halo, and in agent benchmarks it beats the M5 Pro MacBook Pro. The index does not yet record Mac prompt speeds.
Is AMD Strix Halo good for running local LLMs?
It is the cheapest way to get 128 GB of GPU-addressable memory, and it generates text about as fast as a DGX Spark: 56.6 tok/s on gpt-oss-120b with llama.cpp on Vulkan, according to Donato Capitella's published runs. Its weak spots are prompt processing, which is a quarter to a third of the Spark's on an empty context, and a software stack where Vulkan is still steadier than ROCm.
How much does a DGX Spark cost in November 2026?
NVIDIA launched the 128 GB Founders Edition at $3,999 in October 2025 and raised it to $4,699 in February 2026. The Register reported $6,950 on 2 October 2026, though NVIDIA's own pages showed no price on 18 November. A 64 GB version from OEMs starts at $4,999. Dell's GB10 workstation with 128 GB and 2 TB was $7,901 on Dell's US store on 18 November 2026.
What is the cheapest machine that runs gpt-oss-120b?
A Mac mini M5 Pro with 64 GB, at $2,699 from Apple: it holds gpt-oss-120b with about 2 GB to spare, which leaves little room for context. A 128 GB Strix Halo mini PC is the cheapest 128 GB machine, but no longer under $3,000: the Framework Desktop rose from $2,459 in January 2026 to $3,449 by September and was out of stock on 18 November 2026.
Can a DGX Spark or a Strix Halo box run models bigger than 128 GB?
Not on its own. Two DGX Sparks connect over one QSFP cable for 256 GB, and four need a 200 GbE switch. Strix Halo stops at 128 GB; AMD's Ryzen AI Max+ PRO 495 successor goes to 192 GB. A single Mac Studio M5 Ultra is sold with 256 or 512 GB.
Does LLMCheck benchmark the DGX Spark or Strix Halo?
No. LLMCheck runs no tests of its own, and its index covers Apple Silicon Macs only. The DGX Spark and Strix Halo figures on this page are quoted from the published sources linked beside them, with the runtime and date, and they never enter the index or its rankings.