Skip to main contentThe Local Bench

Mac Studio M5 Ultra (base)

Apple · Apple unified · as of 2026-10-12

Specs sourced · published reference runs · catalogue speeds mostly estimated
Mac Studio

  • Data checked: 2026-10-12. Review interval shown below. Prices need review after 30 days; specifications and benchmark collections after 90 days. Dates describe the recorded check, not a live retailer quote.

    • Verified Checked against the primary source (for example the model’s config.json on Hugging Face).
    • Sourced Taken from a cited source; not yet independently checked.
    • Estimate Calculated with the published formulas; not a measurement.
    • Sample Placeholder data for layout; not a real figure.
    • Not yet measured No public, citable score found for this benchmark yet. We never fill the gap with a guess.

    Verified metadata is not a measured performance result. Only cited runs matching model, device, quant and recorded context can replace speed estimates. Estimates assume one stream and F16 KV cache, with a ±25% uncertainty band. A published run applies only to its recorded runtime and conditions; missing settings are unknown.

    • llama-2-7b · mac-studio-m5-ultra
      llama.cpp (Metal) · Q4_0 · tg128
      Recorded 2026-10-12 · historical source record · repeat count unknownSourced
      Llama 2 7B Q4_0, llama-bench tg128, measured on the M5 Ultra 80-core GPU option (1.2 TB/s), commit c1d0e7a. The base 64-core model may be slightly slower. Run source

    Evidence labels · Freshness policy · Calculation methodology

96GB
memory, 72.0 GB usableSourced
1,200GB/s
memory bandwidthSourced
$5,499
price, USDSourced
10of 10
catalogue models fit at 8k

1.2 TB/s of unified bandwidth: the fastest way to run 70B dense models at home.

Decode-speed ceiling for a model is 1,200 GB/s ÷ GB read per token. For a 16.6 GB file that is 73 tok/s; real runtimes reach roughly 45% of it. Why bandwidth is the limit.

See every model on this device

Specifications

Memory
96 GBSourcedBase configuration, apple.com, checked 2026-10-12. Source
Bandwidth
1,200 GB/sSourced1.2 TB/s, apple.com, checked 2026-10-12. Source
FP16 dense compute
80 TFLOPSSampleSample FP16 dense compute; pending vendor-sheet citation.
Max continuous power (system)
480 WSourcedMaximum continuous power of the whole Mac Studio (apple.com); Apple publishes no chip TDP. Source
Load power at wall
240 WSampleSample wall power under load.
Price, USD
$5,499SourcedApple US store, base M5 Ultra configuration. Checked 2026-10-12. Source
Price, CAD
CA$7,799SourcedApple Canada store, base M5 Ultra configuration. Checked 2026-10-12. Source

Models that run

Default quant, single stream, 8k context. Fastest first; estimates are ±25% until measured.

Model
Quant
Memory
Verdict
Decode, tok/s
Max context
gpt-oss-20bMXFP412.9 GBFits235.9Estimate128k
Qwen3 30B-A3BQ4_K_M19.9 GBFits192.1Estimate32k
gpt-oss-120bMXFP464.3 GBFits175.7Estimate128k
Llama 3.1 8B InstructQ4_K_M6.6 GBFits90.1Estimate128k
Qwen3 8BQ4_K_M6.8 GBFits86.6Estimate32k
Qwen3 14BQ4_K_M10.9 GBFits52.2Estimate32k
Mistral Small 3.2 24BQ4_K_M16.2 GBFits34.5Estimate128k
Gemma 3 27B ITQ4_K_M18.2 GBFits30.6Estimate128k
Qwen3 32BQ4_K_M22.5 GBFits24.6Estimate32k
Llama 3.3 70B InstructQ4_K_M45.8 GBFits11.9Estimate64k

Published reference runs

Third-party llama.cpp measurements on this device, with conditions and source. Llama 2 7B Q4_0 runs are the community calibration test; they are not catalogue models.

Model
Quant
Date
Runtime · test
Decode, tok/s
Conditions
llama-2-7bQ4_02026-10-12llama.cpp (Metal) · tg128179.1SourcedLlama 2 7B Q4_0, llama-bench tg128, measured on the M5 Ultra 80-core GPU option (1.2 TB/s), commit c1d0e7a. The base 64-core model may be slightly slower. Source

Usable memory: discrete VRAM − 0.5 GB reserve; unified RAM × 0.75; fits when total ≤ 90% of usable (concept brief §7.3). Markdown · JSON