Mac Studio M3 Ultra · 512 GB
The folk hero of local AI — 671B-parameter models on a silent desk box. Discontinued new; the refurb market's hottest ticket.
Price
was $9,499 / ≈£9,699
Discontinued — refurb and second-hand only; expect volatility
verified 2026-07 · supply & lease options in every proposal
Apple- 512 GB
- Unified memory
- 819 GB/s
- Memory bandwidth
- ~18 tok/s
- DeepSeek 671B @ Q4
- <300 W
- Whole-box draw
one coherent pool
frontier weights at usable speed
verified, MLX
silent under load
This is the machine that proved the category: 512 GB of unified memory at 819 GB/s running DeepSeek R1 671B at 4-bit (~18 tok/s) and Kimi K2-class trillion-parameter weights at ~3.5-bit — under 300 W, in silence, at a price no GPU system could touch.
Apple withdrew the configuration in March 2026 as the memory shortage bit (the $4,000 memory upsell simply stopped being offered). Units circulate refurbished and second-hand — launch pricing was $9,499 / ≈£9,699 — and four of them clustered over Thunderbolt 5 RDMA famously serve 1T-parameter models at ~25 tok/s for the price of one GB300 Station's deposit.
We list it because clients ask for it, the second-hand supply is real, and the M5 Ultra successor (tested to 768 GB, expected late 2026) will reopen this exact play with new silicon. If you find clean stock at sane money, it remains the cheapest 400 GB-class inference node ever sold.
vs its siblings: The value anomaly of the whole catalogue — while stock lasts.
Memory, to scale
512 GB model-visible · bandwidth is the speed limit
Unified memory
512 GB · 819 GB/s
LPDDR5x
For scale
DGX Spark — 128 GB @ 273 GB/s
RTX PRO 6000 — 96 GB @ 1.79 TB/s
DGX Station GB300 — 748 GB coherent
4× 512 GB Studios over TB5 RDMA ≈ 2 TB pooled — Kimi-class 1T models at ~25 tok/s (verified community builds).
What it actually runs
Declared from research and benchmarks, not computed marketing — tokens-per-second figures are cited where a real measurement exists.
- DeepSeek 671B-classQ4 (~404 GB)fits~18 tok/s generation — the proof-point benchmark
- GLM-5.24-bit (~390 GB)fitsthe current open-weight leader, on a desk
- Qwen3.5 397BInt4 officialwith headroom~205 GB — real headroom for context
- MiniMax M3INT4with headroom~220 GB — multimodal frontier class
- Kimi K2.6 (1T)INT4 (594 GB)2+ units linked2× Studios over TB5 RDMA (~15 tok/s)
- gpt-oss-120bMXFP4with headroom~60 tok/s — with 450 GB to spare
The full sheet
Compute
- Chip
- Apple M3 Ultra — 32-core CPU, 80-core GPU (required for 512 GB)
- Neural Engine
- 32-core
Memory
- Unified memory
- 512 GB @ 819 GB/s
- Model capacity
- DeepSeek 671B Q4 (~404 GB weights, ~448 GB wired) — verified
Storage & I/O
- SSD
- 1 TB base, to 16 TB
- Ports
- 6× Thunderbolt 5 · 10 GbE
Status
- Availability
- Discontinued (Mar 2026) — refurb / second-hand / residual channel stock
- Successor
- M5 Ultra Mac Studio expected late 2026, tested to 768 GB
Where it shines
- The cheapest 400 GB-class inference node ever sold
- Silent, <300 W — no facilities conversation required
- Proven on the biggest open weights (671B verified, 1T at low-bit)
- TB5 RDMA clustering scales it to trillion-parameter class
The trade-offs
- Discontinued — supply is refurb/second-hand with price volatility
- Prompt processing is the weakness: ~15 min to ingest an 8k prompt in llama.cpp (MLX is ~4–5× better but still trails NVIDIA)
- No CUDA ecosystem; MLX/llama.cpp serving only
- No warranty story on grey-market units — buy through refurb channels
Buy this box for
Understanding Apple Mac Studio
Apple M3 Ultra
Apple silicon's unified-memory architecture gives the GPU the entire RAM pool at 819 GB/s — 3× a DGX Spark's bandwidth. The M3 Ultra's 512 GB configuration became the folk hero of local AI, running DeepSeek R1 671B at 4-bit (~18 tok/s) and even Kimi K2-class trillion-parameter models at ~3.5-bit, silently, under 300 W.
The honest caveat for buyers today: the memory shortage killed the big configurations. Apple withdrew the 512 GB option in March 2026 and the 256 GB option by May — new orders cap at 96 GB (£5,299 since the June 2026 price rise). The 512 GB machines circulate refurbished and second-hand, and the M5 Ultra refresh expected late 2026 has been tested to 768 GB. Clustering is the workaround that works now: macOS 26.2 added RDMA over Thunderbolt 5, and four M3 Ultras pool ~2 TB to serve trillion-parameter models at ~25 tok/s.
The other trade-off is prompt processing: no CUDA, so you serve through MLX or llama.cpp, and long-context ingestion is far slower than NVIDIA silicon even though generation speed is strong. Right machine for chat and agent workloads; wrong machine for 100k-token document dumps.
Sources & verification
Specifications and prices verified 2026-07 against the sources below. The memory shortage is repricing this market monthly — we re-verify at quote.
Compare against the rest of the catalogue, or have us spec it against your workloads.
Back to catalogueSpec the Mac Studio M3 Ultra · 512 GB for your workload
We'll size it against your real token volumes, quote supply and lease options, and stand up the models with evals that prove they do the job — or tell you honestly that a different box (or the API) wins.