M5 Ultra Mac Studio Makes Local AI Models Faster to Respond

What Apple shipped

Apple's M5 Ultra arrived on August 25, 2026, with a refreshed Mac Studio. Spec-sheet coverage has focused on the 80-core GPU and 8K video. The more lasting change is in how the chip runs large language models. At the top configuration, the chip has an up-to-36-core CPU, an up-to-80-core GPU and 1.2TB/s of unified memory bandwidth. Apple says that bandwidth is 50 percent higher than the M3 Ultra's.2 Memory goes up to 512GB, and Apple describes the machine as able to run enormous LLMs entirely on device.4 Pre-orders opened on launch day and the Mac Studio went on sale September 22. The 512GB configuration, which most AI teams want, slips to late October.6

Pricing puts it in professional territory. The Mac Studio with M5 Ultra starts at $5,499 with 96GB of memory and 1TB of storage.3 Moving from the base 30-core CPU and 64-core GPU to the 36-core CPU and 80-core GPU adds $1,300. That makes the full-GPU configuration $6,799 before any memory or storage upgrades.5

The core count didn't change. What's inside each core did

The 80-core figure is not new. The M3 Ultra also topped out at 80 GPU cores, so the AI gains come from changes to the architecture, not from adding cores.4 The new part is a Neural Accelerator in every GPU core: dedicated matrix-multiply hardware, the kind of math that dominates transformer inference.6 The M5 Max introduced these accelerators in the MacBook Pro. The M5 Ultra is the first Ultra chip to have them.11

Apple's own numbers don't quite match each other. The M6 and M5 Ultra newsroom release claims up to 4.5x the M3 Ultra's peak GPU compute for AI, and more than 6x the M1 Ultra's.2 The Mac Studio announcement uses a different multiple: up to 4.3x the M3 Ultra.20 Outlets have repeated both figures, sometimes in the same article.412 For actual workloads, Apple says LLM prompt processing in LM Studio is up to 4x faster than on the M3 Ultra and up to 9.8x faster than on the M1 Ultra.20 Those are Apple's tests on preproduction systems, which compared 256GB configurations of the M5 Ultra and M3 Ultra.2

The gap between 4.3x and 4.5x matters less than the type of claim. Both are peak compute figures. As one analysis put it, the headline multiple describes the best case, not what a typical workload sees.6

Prompt processing vs. token generation

Local inference has two phases with different bottlenecks. The independent coverage agrees that the M5 Ultra speeds them up by very different amounts.

In the first phase, prefill, the model reads the whole prompt in one large matrix multiplication. This phase is limited by compute, which is exactly the work the Neural Accelerators target.12 In Apple's MLX tests on the base M5 chip, time-to-first-token was 3.3x to 4.06x faster than on the M4. Each token after that was only 1.19x to 1.27x faster, because generation was still held back by a smaller bandwidth increase.17 Ars Technica's reviewer made the same point in practical terms: the GPU accelerators greatly cut the wait before a model starts answering, which makes any local model feel more responsive.16

The second phase, decoding, is limited by memory bandwidth. The key figure there is the jump from roughly 819GB/s to 1.2TB/s. Bandwidth had stayed at about 819GB/s across three generations of Ultra chips.4 One analysis argues that bandwidth translates almost directly into generation speed when decoding is memory-bound. It notes the M5 Ultra has about double the M5 Max's 614GB/s.6

The split explains the agentic workflows reviewers kept highlighting. Agents resend long contexts on every turn, so faster prefill pays off over and over.

What reviewers measured

The hands-on numbers are strong, though mostly they come from a few mixture-of-experts (MoE) models. MacStories ran Qwen3.8-Flash-Next and got more than 100 tokens per second on short prompts. Generation stayed at 60 to 85 tokens per second with 64K to 256K tokens of context, enough to make tool-calling agents feel fluid.12 Context Studios measured a median of 108 tokens per second on short prompts with the same model, 54 percent more than the M3 Ultra's 70. It also reported prefill roughly 2.5 times faster than the previous generation.13

That 2.5x real-world prefill gain sits well below Apple's 4x prompt-processing claim. It's the clearest gap between marketing and measurement so far. The chip isn't underdelivering, but buyers should expect gains that vary by model.

Ars Technica came at it from a different angle. Its reviewer ran the open-weight Qwen 3.8 27B model locally in LM Studio Bionic for coding and called it the most fun he'd had with new technology in a long time.16 Kingy.ai was more cautious. As of mid-September, it judged the benchmark picture unfinished, noting that independent results were still scarce and that Geekbench 7 had only a single upload.5

Why 512GB and MoE models go together

Several analysts reached the same conclusion: sparse MoE models are the M5 Ultra's best use case. A large MoE model needs a lot of memory to hold all its weights, but each token only reads the active parameters. That plays to a machine with huge capacity and fast bandwidth.11

The 512GB tier pushes the limit into territory that used to require multiple GPUs. Pinggy's fit analysis reports that a 4-bit MLX build of GLM-5.2, a 743-billion-parameter model with 40 billion active, takes 418GB. That fits on one machine at the edge, with a raised memory limit and modest context. Models like Kimi K3 and Qwen3.8-Max, at 2.4 to 2.8 trillion total parameters, still need clusters on any hardware.17 ModelFit estimates the 512GB configuration runs Llama 4 Maverick 400B at about 12 tokens per second and DeepSeek-R1 671B at about 6. These are modeled figures, not measurements.15

The buying advice from Context Studios is sensible. It calls 256GB, starting at $9,499, the sweet spot for agentic work with large MoE models, and recommends 512GB only when a specific model won't fit in 256GB.13 For buyers mainly running models of around 70 billion parameters, it says a 128GB M5 Max Studio at roughly half the price is often the better buy.13

Clustering and software

Apple is also pushing beyond a single box. Owners can pool memory across several Mac Studios, and Apple claims a four-machine cluster delivers 3x the inference throughput of one system.11 Apple's test footnote describes a four-node full-mesh setup over Thunderbolt 5 with RDMA, running a 72-billion-parameter dense model through MLX.19 On the software side, macOS 27 adds Core AI, a framework for deploying full-scale LLMs locally that works alongside MLX.11 In practice, multi-Mac clusters are the only realistic way to run trillion-parameter-class models on Apple hardware.17

How it compares with Nvidia

On prefill, Nvidia keeps an advantage. MacStories found that the Neural Accelerators narrow the gap with the RTX 5090's Tensor Cores but don't close it.12 Community tests collected by Context Studios show the same pattern. On GLM-5.3-Flash, one M5 Ultra posted 28 tokens per second decode and 1,016 prefill. A pair of DGX Spark units reached 39 and 1,700.13 Each DGX Spark has 128GB at 273GB/s, so the M5 Ultra leads clearly on memory capacity and bandwidth per box. Nvidia's advantage is its purpose-built CUDA software ecosystem.5

One caution on older coverage. A pre-launch analysis published in April assumed the M5 Ultra would have 192GB and roughly 800GB/s. Apple's shipping specs are far higher, so pre-announcement benchmark tables should be treated as speculation.18

Bottom line

The headline claim that the M5 Ultra will power through AI models holds up, with a few qualifications. The 4.3x to 4.5x multiples describe peak compute. The real story is faster prompt processing combined with about 46 percent more bandwidth, attached to a memory pool large enough to hold models with hundreds of billions of parameters. The M5 Ultra is not a CUDA replacement and won't match Nvidia on raw prefill. It is the strongest case yet for running frontier-scale open models quietly on a desk. The 512GB configuration, the one that best shows that case, doesn't ship until late October, and independent testing of it is still missing.