Mac Studio M5 Ultra for Local LLMs: Prices, 256 GB vs 512 GB, Speed and What Fits

The Mac Studio M5 Ultra is the fastest Mac for local LLMs: 1.2 TB/s of memory bandwidth with 96 GB from $5,499, 256 GB from $9,499 or 512 GB for $22,299. The 96 GB model runs gpt-oss-120b, 256 GB holds 4-bit builds of GLM-5.3-Flash and Qwen3.8-Flash-Next, and only 512 GB holds GLM-5.3 or DeepSeek-V4.1-Flash in memory, as 3- to 4-bit community builds. By our bandwidth estimate, gpt-oss-120b tops out near 144 tokens per second for one user.

Apple announced the Mac Studio with M5 Max and M5 Ultra on 25 August 2026, began deliveries on 22 September and put the 512 GB option on sale on 28 October.

Prices come from Apple’s US store, specifications from Apple’s pages and newsroom, and model sizes from Hugging Face, all checked on 4 November 2026. Each speed is either a credited measurement or an estimate from the method behind our tokens-per-second calculator. Picks for every other Mac are in our guide to local models across the Mac range.

Mac Studio M5 Ultra price, release date and configurations

The store sells 256 GB with either Ultra chip, but 512 GB only with the 36-core CPU and 80-core GPU: $15,500 over that chip’s 96 GB price, or $11,500 over 256 GB. Storage goes up to 16 TB for $7,500 more, and education pricing starts at $5,099.

Apple builds the M5 Ultra from two M5 Max chips joined by UltraFusion. Each of its 64 or 80 GPU cores includes a Neural Accelerator for matrix maths, beside a 32-core Neural Engine, and both chip options run memory at 1.2 TB/s, which Apple puts at 50 percent above the M3 Ultra. The machine has six Thunderbolt 5 ports, 10 Gb Ethernet and a 480 W maximum continuous power rating.

Memory and bandwidth: what the M5 Ultra changes for local models

Capacity decides which models load; bandwidth decides how fast they answer, because each generated token reads the active weights once. The M5 Ultra’s 1.2 TB/s is about twice a 40-core M5 Max’s 614 GB/s, 4.4 times DGX Spark’s 273 GB/s and two-thirds of an RTX PRO 6000’s 1,792 GB/s. Our desk-side system comparison covers the NVIDIA and AMD boxes.

Sizing follows our VRAM calculator: weights take parameters × bits ÷ 8, which for a published build is its download size; the KV cache adds fixed bytes per token of context; and a fifth of unified memory stays free for macOS and the runtime. That leaves 76.8 GB usable on 96 GB, 204.8 GB on 256 GB and 409.6 GB on 512 GB. MLX LM documents raising the GPU’s wired memory limit with sudo sysctl iogpu.wired_limit_mb=N, which helps only models that already fit.

Reading a prompt is limited by arithmetic rather than bandwidth, so it scales with GPU cores. The $1,300 step to the 80-core GPU therefore buys 25 percent more cores for prompt processing at the same bandwidth: worth it for coding agents and retrieval, which send long prompts every turn, and less so for short chats.

Which models fit in 96 GB, 256 GB and 512 GB

Sizes are Hugging Face file listings, mostly community conversions, checked 4 November 2026. Cache figures use each model’s attention layout at 16 bits: Qwen3.8-27B caches 64 KiB per token in 16 of its 64 layers, Qwen3.8-Flash-Next about 27 KB in 12 of 48. Our Qwen 3.8 sizing guide, GLM-5.3 deployment guide and gpt-oss setup guide go deeper.

Check licences too. gpt-oss and Qwen3.8-27B use Apache 2.0, and GLM-5.3-Flash and DeepSeek-V4.1-Flash use MIT. Under the Qwen Community License 1.0, products built on Qwen3.8-Flash-Next with over 100 million monthly users or US$20 million in monthly revenue must display its name, and a group running a Model-as-a-Service or AI Work Assistant business needs a separate licence before commercial use, unless that use stays internal. GLM-5.3’s own licence adds a security review for very large model-as-a-service groups.

Near misses and models that do not fit one Mac Studio:

• GLM-5.3-Flash in MLX 4-bit (204.0 GB) leaves under 1 GB on 256 GB. MacStories saw a mixed 4- and 8-bit build run out of memory on a 256 GB M5 Ultra during its 128K and 256K-token tests.
• GLM-5.3 is 755.6 GB in FP8, and mlx-community’s 4-bit build, 418.3 GB, exceeds even the 512 GB budget.
• DeepSeek-V4.1-Flash ships at 510.3 GB and its MLX 4-bit build is 552.9 GB, so one Mac holds only 3-bit and smaller builds entirely in memory; our DeepSeek-V4.1-Flash local guide covers reading its Engram tables from SSD instead.
• Kimi K3 is about 1.56 TB as published, and Unsloth’s smallest GGUF, 466.4 GB, still exceeds the 409.6 GB budget. Our Kimi K3 deployment guide covers the multi-node route.
• Qwen3.8-2.4T-A95B takes 2,496 GB in FP8.

M5 Ultra tokens per second: estimates and measurements

Our calculator divides usable bandwidth, at 60 percent of peak, by the bytes each token reads, including KV cache for a 2,048-token prompt and 512-token reply. Results are one-user estimates, not benchmarks.

Estimated tokens per second; our GLM-5.3 guide’s figure of about 64 for GLM-5.3-Flash assumes the larger UD-Q4_K_XL build.

The llama.cpp results imply 57 to 83 percent of peak bandwidth, depending on precision, so our 60 percent is cautious for dense models. Mixture-of-experts models with few active parameters fall further short: Qwen3.8-Flash-Next activates 6 billion, for an estimate near 190 tokens per second against the 90 to 110 MacStories measured for prose. Concurrency divides speed, too: four simultaneous gpt-oss-120b requests get about 68 tokens per second each by our method.

Prompt processing: Apple’s claims and what testers saw

Apple’s main AI claim for the M5 Ultra concerns prompts: time to first token in LM Studio up to 4 times faster than on M3 Ultra and 9.8 times faster than on M1 Ultra, measured with an 8K-token prompt, a 14-billion-parameter 4-bit model and LM Studio v0.4.19+2 on a preproduction machine in July 2026. Its chip announcement claims up to 4.5 times the M3 Ultra’s peak GPU compute for AI; the Mac Studio release says 4.3 times. Apple publishes no TFLOPS figure, so our calculator cannot estimate a Mac’s time to first token.

Independent figures agree on the direction. In llama.cpp’s table the M5 Ultra reads prompts 3.3 times as fast as an M2 Ultra, and MacStories measured Qwen3.8-Flash-Next reading 2,500 to 2,900 tokens per second against about 1,100 on an M3 Ultra, so a 256K-token prompt reached its first token in 104 seconds instead of 246. NVIDIA still reads prompts faster, as the RTX 5090 result above shows.

Mac Studio M5 Ultra and MLX, LM Studio, llama.cpp and Ollama

All four common runtimes support the M5 Ultra as of 4 November 2026:

• MLX and MLX LM. Apple’s open-source framework runs models on the GPU through Metal. MLX 0.32.3 (29 September) added M5 Ultra matrix-multiplication tunings, and MLX LM 0.32.2 (16 October) supports Kimi K3 and DeepSeek-V4.1. Its documentation says the built-in HTTP server is not recommended for production because it implements only basic security checks.
• LM Studio. Loads GGUF and MLX models and is the tool behind Apple’s prompt claim; version 0.4.28 (31 October) is current.
• llama.cpp. Apple silicon is a first-class Metal target, and the same GGUF file later runs on Linux servers. Build b12102 shipped on 4 November; DeepSeek-V4.1-Flash support was merged on 25 October.
• Ollama. Since version 0.40.0 (25 September), it runs architectures its MLX engine supports on MLX by default on Apple silicon, starting with Qwen 3.8, Qwen 3.6, Qwen 3.5 and Gemma 4; 0.40.14 (3 November) is current. Its local API requires no authentication, so add an access layer before colleagues connect.

A first MLX run on the 96 GB model:

pip install -U mlx-lm
mlx_lm.generate --model mlx-community/gpt-oss-120b-MXFP4-Q8 --prompt "Summarise the main risks in this supplier contract: ..." --max-tokens 400

Clustering Mac Studios over Thunderbolt 5

Apple supports linking Mac Studios over Thunderbolt 5 with RDMA, remote direct memory access, so one model can span their memory. It claims up to three times faster inference from four systems than from one, measured on a 32K-token summarization with a 72-billion-parameter dense model in MLX LM across four 512 GB machines cabled as a full mesh.

MLX’s documentation says RDMA over Thunderbolt arrived in macOS 26.2, is enabled from macOS Recovery with rdma_ctl enable, and, for its JACCL backend, needs every Mac cabled directly to every other, which six ports allow for four machines. exo, an Apache 2.0 project that Apple’s page names for clustering, also needs macOS 26.2 and claims up to 1.8 times the speed on two machines and 3.2 times on four.

Two 512 GB machines ($44,598) give about 819 GB under our rule, enough for mlx-community’s 4-bit GLM-5.3 or Unsloth’s 711.1 GB 2-bit Kimi K3. Four ($89,196) give about 1,638 GB, just above Kimi K3’s 1,561 GB checkpoint. Each link carries up to 120 Gb/s, far below the 1.2 TB/s inside each machine, so measure before trusting vendor speed-ups.

Mac Studio M5 Ultra vs RTX PRO 6000

For a model that fits in 96 GB, the NVIDIA card is faster: about 1.5 times the bandwidth lifts gpt-oss-120b’s estimated ceiling from 144 to 215 tokens per second, its published compute favours prompt processing, and it runs the CUDA serving stack of production GPU servers. Our serving-card comparison of the RTX PRO 6000 with H100 and H200 goes further.

The Mac wins on capacity per dollar and per watt. $9,499 buys 256 GB of model memory in a 480 W computer; matching it takes three RTX PRO 6000 cards, $48,000 and 1,800 W of board power before the workstation.

M5 Max vs M5 Ultra for LLM work

Double the bandwidth doubles our estimates, 144 against 74 tokens per second for gpt-oss-120b, though the small llama.cpp test measured nearer 1.5 times. One case favours the Max: at $5,099 with 128 GB it holds more than the $5,499 Ultra with 96 GB, so a 77 to 102 GB model such as Qwen3.8-Flash-Next at IQ4_XS fits only the cheaper machine. Choose the Max for capacity on a budget, the Ultra for speed, long prompts or more than 128 GB.

Which Mac Studio M5 Ultra to buy: 96 GB, 256 GB or 512 GB

• 96 GB, $5,499. The lowest price per unit of bandwidth in the Mac Studio line. It holds gpt-oss-120b with long context, or dense 27B to 32B models at 8-bit, for one developer.
• 256 GB, $9,499 or $10,799. The tier for the strongest models most teams can run on a desk: 4-bit Qwen3.8-Flash-Next, GLM-5.3-Flash or DeepSeek-V4-Flash, with cache for several sessions. Add the 80-core GPU if long prompts dominate.
• 512 GB, $22,299. Only for evaluating GLM-5.3 or DeepSeek-V4.1-Flash in 3- to 4-bit community builds before buying GPU servers; the $11,500 over 256 GB buys capacity alone.
• Not a Mac Studio. For shared serving, CUDA parity or prompt-heavy work, look at GPU servers. Under about 30 billion parameters, one or two 32 GB cards such as AMD’s Radeon AI PRO R9700, $1,299 each at launch, cost less.

How VDF AI fits

A Mac Studio can host the model, but VDF AI’s platform does not run on macOS in production. The self-hosting platform requirements treat Linux as the production platform; macOS runs through Docker Desktop for evaluation and development.

The two meet in a pilot. VDF AI Router can register the Mac’s Ollama endpoint as a custom on-premises deployment, listed beside whichever cloud models policy allows, and in air-gap mode it sends requests to local models alone. Governed agents, multi-agent orchestration with Networks, private RAG, role-based access control and audit logs stay in the platform, so moving the model to Linux GPU servers later changes only the router’s deployment entry. VDF Code, the governed coding assistant, also runs on infrastructure you control.