M5 Ultra 256GB: Real-World Performance & 671B Model Local Inference Tested

Conclusion Upfront

For 4K video editing, coding, or dozens of browser tabs, the M5 Ultra 256GB is severe overkill. However, if you need local large-model inference, massive 3D/video rendering, or a "desktop supercomputer," it is currently the only Apple Silicon machine that can fit a 671B-parameter model (e.g., DeepSeek-R1) entirely in memory without disk swapping, delivering a usable 10–15 tokens/sec.

Specifications: What Changed

CPU: 32 cores (24 performance + 8 efficiency) vs. M2 Ultra's 24 cores (16P+8E) — a 33% increase.

GPU: 80 cores vs. 76 cores — a modest bump.

Neural Engine: 32 cores, ~40 TOPS vs. 16 cores, 31.6 TOPS.

Unified Memory: Up to 256GB vs. 192GB max.

Memory Bandwidth: ~1024 GB/s vs. 800 GB/s — first Apple chip to break 1 TB/s.

Transistors: ~190 billion vs. 134 billion.

In short: CPU up 1/3, GPU slightly improved, but the real gap is the 256GB memory and 1 TB/s bandwidth.

Benchmarks: Multi-Core Nears 40k

Leaked Geekbench 6 multi-core scores place M5 Ultra in the 38,000–40,000 range, roughly 35–40% above M2 Ultra's ~28,000. That puts it ahead of many mainstream flagship desktop CPUs (~32,000) and far above high-end gaming laptops (~18,000), while consuming less than half the power of comparable HEDT platforms. Single-core gains are modest; daily app responsiveness feels similar to M4 series.

Why 256GB Unified Memory Matters

Model size is fundamentally limited by VRAM/RAM. The article references a cheat sheet (previously shared) for quick lookup:

70B model at Q4 quantization ≈ 42GB — fits in 64GB RAM.

DeepSeek-R1 671B (full-weight) even at Q2/Q3 quantization requires >200GB.

Only machines with 192GB+ can hold the entire 671B model in memory without swapping to disk.

M5 Ultra's 256GB yields ~250GB usable space, enabling a "cloud-grade" model on a desktop. Inference speed reaches 10–15 tokens/sec — usable for interactive chat without internet.

Who Should Buy (and Who Shouldn't)

Worth it: Local LLM researchers, AI application developers, 3D/video render farms, workloads needing single-machine large memory.

Skip: General content creation, programming, office work — M4 Max (64–128GB) is plenty; savings could buy another monitor.

Caveat: The 256GB configuration costs nearly 7× the base model. The price/performance ratio only makes sense if you genuinely saturate that memory.

Final Takeaway

The M5 Ultra 256GB is not for everyone, but it represents the current ceiling for desktop "local AI" on Apple Silicon. If you're still deciding how much memory you need or which models you can run, consult the referenced cheat sheet for a per-tier breakdown from 8GB to 256GB.