M5 Ultra vs M3 Ultra vs M5 Max for Local LLMs: The Measured Numbers

The three chips

Sources: Apple’s Mac Studio specs and 25 August announcement; the M3 Ultra’s 819 GB/s is from Apple’s MLX research post. The 512 GB M5 Ultra requires the 80-core GPU, and Apple had not announced its price as of 18 October. The 36 GB M5 Max uses the binned 460 GB/s chip, so for local models start at 64 GB.

Decode tracks bandwidth

Generating each token means reading every active weight from memory once, so decode speed follows memory bandwidth. The M5 Ultra has 1.50× the M3 Ultra’s. Every plain-decode pair in the index that comes from a single source lands close to that:

Eight pairs from four sources land between 1.33× and 1.58×, around the 1.50× bandwidth ratio; Apple’s own summary is 1.6×. BGR reported rounded figures, which accounts for some of the spread at the low end. The M5 generation’s faster GPU barely shows here, because the GPU spends most of each token waiting on memory.

Where the M5 Max fits

With 614 GB/s, the M5 Max has half the M5 Ultra’s bandwidth and three-quarters of the M3 Ultra’s. The index’s figure for each chip:

Small mixture-of-experts models are where the M5 Max shines. A model with 3B active parameters spends much of each token on fixed per-token work rather than on reading weights, so bandwidth matters less and the newer GPU more. In MLX at 4-bit, a community oMLX run measured Qwen 3.6-35B-A3B at community125.8 tok/s on an M5 Max (1K context, run), against community106.4 tok/s on an M3 Ultra and community153.4 tok/s on an M5 Ultra in NetworkChuck’s runs. The harnesses differ, but the engine and bit width match: the M5 Max beats the M3 Ultra and reaches 82% of the M5 Ultra.

As active parameters grow, the gap returns to the bandwidth ratio. Qwen 3.5 122B-A10B (10B active) was measured at community42.9 tok/s on an M5 Max in llama.cpp (Q4_K_M, baem1n) against community87.7 tok/s on an M5 Ultra in MLX. Part of that gap is the engine: an arXiv study on one M4 Max measured mlx-lm 19% ahead of llama.cpp on Qwen3-30B-A3B at 4-bit (107.4 against 89.9 tok/s, published). Across its models the gap ran from −14% to +27%. Dense models scale with bandwidth all the way: tbreak measured Llama 3.3 70B at sourced13.1 tok/s on an M5 Max (4-bit MLX, review), and for GPT-oss 120B sourced75.65 tok/s in Ollama. The index still estimates the M5 Max for dense Qwen3.8-27B, at 4-bit; BGR’s review says an M5 Max MacBook Pro performed similarly to the M3 Ultra on it, but published no figure.

Prompt processing is the M5’s real gain

Before a model writes its first token it must read the whole prompt, and that step is limited by compute, not by bandwidth. Apple’s November 2025 MLX post explains why the M5 generation pulls ahead here: its GPU cores carry Neural Accelerators that speed up the matrix work of prefill, while generating later tokens stays bound by memory bandwidth. Between the two Ultras, that shows up as about a 4× gain on dense models and 3× on mixture-of-experts models:

The M5 Max sits between them. A community oMLX run read a 4K prompt for Qwen 3.6-35B-A3B at community4,314 tok/s on an M5 Max — about half the M5 Ultra’s 8,799 and 1.4× the M3 Ultra’s 3,061, as its 40 GPU cores against the Ultra’s 80 would suggest. Different harnesses again, so read the ratios, not the decimals.

This matters more for agents than for chat. A coding agent re-reads a long context on every step, so prompt reading is most of its waiting. At a 32K-token prompt, NetworkChuck’s server runs of Qwen3.8-27B read at community388 tok/s on the M3 Ultra and community1,526 tok/s on the M5 Ultra — a long-prompt workload, outside the index’s plain-decode rows. Power rises too: Apple measured peaks of about 200 W for the M3 Ultra and 300 W for the M5 Ultra at a 65,536-token context.

Which one to buy

• The M5 Max, if your models fit in 128 GB. The popular small mixture-of-experts models run at 100+ tok/s on it, at or above the M3 Ultra, and on those models it reads prompts faster than the M3 Ultra too. The index’s 128 GB pick, Qwen3.8-Flash-Next, runs at community44 tok/s on it. Choose 64 GB or more: the 36 GB base has the 460 GB/s chip. Best LLMs for the 128 GB M5 Max.
• The M3 Ultra, if you need 512 GB before the M5 Ultra version ships. Decode runs at about two-thirds of the M5 Ultra’s speed, and prompt reading at a quarter to a third of it. That is fine for chat with short prompts and slow for long-context agents. Best LLMs for the 512 GB M3 Ultra.
• The M5 Ultra, for long contexts, dense models above about 30B, or anything over 128 GB. A 256 GB M5 Ultra runs the leaderboard’s #1, MiMo-V2.6-Flash, at estimated~79 tok/s. The 512 GB model opens the table below. Best LLMs for the 512 GB M5 Ultra.

Two or four Macs over Thunderbolt 5

Apple’s 15 October post also documents MLX across several Mac Studios, linked by RDMA over Thunderbolt 5. It needs macOS 26.2 or later, MLX 0.30.1 or later and MLX’s jaccl backend. Muse Spark Open at 4-bit NVFP4 takes about 493 GB, which Apple says is too much for one 512 GB Mac once the system and the context need room. Apple reports 24.99 tok/s on two M5 Ultras and 36.73 on four, with tensor parallelism. These are vendor-reported multi-Mac figures, which the index does not list as rows.

Only tensor parallelism speeds up decode. Apple’s Qwen3.8-27B BF16 figure rises from 19.69 tok/s on one M5 Ultra to 33.27 on two and 50.89 on four. Pipeline parallelism across two Macs gave 19.44, which adds memory, not speed. The tooling is young, too. Two MLX bugs were open on 18 October: a silent hang when a peer drops out (#3910) and a segfault when a Thunderbolt cable is unplugged mid-run (#4192). A dead-peer fix, #4530, was closed unmerged. If a model fits one 512 GB Mac at 2–3 bits, one Mac is the simpler machine; Muse Spark Open on a Mac walks through that single-Mac path.

What fits in 512 GB

Before the table: a 256 GB Mac Studio already runs the leaderboard’s #1, MiMo-V2.6-Flash (about 172 GB at 4-bit). The models below are the ones that need more, or barely fit 256 GB.

Three notes on reading it. File sizes close to 512 GB also need macOS’s GPU memory limit raised (the iogpu.wired_limit_mb setting), plus room for the context. “Index, M5 Ultra” is the index’s figure for that chip — today an estimate on every row — and the index’s estimates for large mixture-of-experts models have erred in both directions against measured runs (how the estimates compare with measured runs). Kimi K2.7 Code has one community report of about 21 tok/s on a 512 GB M3 Ultra in the Inferencer app, but it does not state its decode mode or the app version, so the index does not list it.

One more model is due. Mistral Large 4 (1T total, 49B active, 52B with embeddings) has a Hugging Face placeholder giving 31 October as the weights date, with no licence named. The index adds it only when the weights are downloadable.

Frequently Asked Questions