More
Submit build

Builds / Apple M5 Pro (unified, Mac mini)

Mac mini M5 Pro 64GB (llama.cpp IQ4_NL): Inkling 2 Mini at 29.95 tok/s short context, 17.39 tok/s at 128k (pratikgx)

Last updated:

Source: huggingface.co β†—πŸ“ How measured: reported by the original source; RigBase re-checked the numbers at that source
GPU(s)
Apple M5 Pro (unified, Mac mini)
VRAM / unified memory (GB)
64
CPU
Apple M5 Pro
Motherboard
not stated
RAM
64 GB unified memory
PSU
not stated
Other
macOS 27.0.1 (26A434), AC power; llama.cpp PR #32694 (official inkling2_loop support), commit d7cb78c91, version 1.0.0-dev (build 12587), Metal, ngl=99, flash-attn on; persistent llama-server with a 133,120-token context slot; official ggml-org Inkling-2-Mini-IQ4_NL.gguf (17,627,940,768 bytes); temperature 0, seed 42, reasoning_effort none; 256 generated tokens, 1 warmup + 3 timed runs, median reported
Reported total cost
not stated
Power
not stated

Measured speed (generation tokens/sec)

ModelQuantEnginetok/sPrompt processing (tok/s)Details
Inkling 2 Miniggml-org IQ4_NL (17.63 GB; 4-bit experts in layers 0-19, Q8_0 elsewhere, loop controller F32)llama.cpp PR #32694 (d7cb78c91, build 12587), Metal29.95491 @ 512-token prompt
est. TTFT @8k β‰ˆ 16.7 s (calculated from measured prefill)
measured TTFT: 1.1 s @ 512-token prompt
context: 512-token prompt, 256 generatedπŸ“ Independent measurement by pratikgx, 20 Nov 2026. Median of 3 timed runs after 1 warmup; largest generation-speed spread at any context was 0.81%. Peak RSS 23.10 GB includes a reserved 128k KV slot, so it is not a tight short-context footprint. Time to first token from the streaming /completion endpoint.
Inkling 2 Miniggml-org IQ4_NL (17.63 GB; 4-bit experts in layers 0-19, Q8_0 elsewhere, loop controller F32)llama.cpp PR #32694 (d7cb78c91, build 12587), Metal28.42358 @ 4,096-token prompt
est. TTFT @8k β‰ˆ 22.9 s (calculated from measured prefill)
measured TTFT: 11.5 s @ 4,096-token prompt
context: 4,096-token prompt, 256 generatedπŸ“ Same machine, runtime and settings as the 512-token row. Median of 3 timed runs. Peak RSS 23.53 GB (includes reserved 128k KV slot).
Inkling 2 Miniggml-org IQ4_NL (17.63 GB; 4-bit experts in layers 0-19, Q8_0 elsewhere, loop controller F32)llama.cpp PR #32694 (d7cb78c91, build 12587), Metal26.74248 @ 16,384-token prompt
est. TTFT @8k β‰ˆ 33.0 s (calculated from measured prefill)
measured TTFT: 66.0 s @ 16,384-token prompt
context: 16,384-token prompt, 256 generatedπŸ“ Same machine, runtime and settings as the 512-token row. Median of 3 timed runs. Peak RSS 24.13 GB (includes reserved 128k KV slot).
Inkling 2 Miniggml-org IQ4_NL (17.63 GB; 4-bit experts in layers 0-19, Q8_0 elsewhere, loop controller F32)llama.cpp PR #32694 (d7cb78c91, build 12587), Metal24.97202 @ 32,768-token prompt
est. TTFT @8k β‰ˆ 40.6 s (calculated from measured prefill)
measured TTFT: 163 s @ 32,768-token prompt
context: 32,768-token prompt, 256 generatedπŸ“ Same machine, runtime and settings as the 512-token row. Median of 3 timed runs. Peak RSS 24.54 GB (includes reserved 128k KV slot).
Inkling 2 Miniggml-org IQ4_NL (17.63 GB; 4-bit experts in layers 0-19, Q8_0 elsewhere, loop controller F32)llama.cpp PR #32694 (d7cb78c91, build 12587), Metal17.3995.29 @ 131,072-token prompt
est. TTFT @8k β‰ˆ 86.0 s (calculated from measured prefill)
measured TTFT: 1,376 s @ 131,072-token prompt
context: 131,072-token prompt, 256 generatedπŸ“ Same machine, runtime and settings as the 512-token row. Median of 3 timed runs. The model stayed in memory. Peak RSS 25.86 GB (includes reserved 128k KV slot). TTFT 1,376 s (22.9 min).

Prompt processing = prefill speed (how fast the prompt is read) as reported by the source, with the prompt length it was measured at; blank = the source does not report it. est. TTFT @8k is calculated from the measured prefill speed (8,192 Γ· prefill tok/s), shown only when prefill was measured on a prompt of 512+ tokens. It is not a measurement and ignores load time and caching. measured TTFT is the source's own time-to-first-token. Why prefill matters for coding assistants.

Independent local measurement by pratikgx, 20 Nov 2026, published as a Hugging Face dataset under CC BY 4.0. RigBase read the dataset card on 21 Nov 2026 and checked each figure against the published table. RigBase did not re-run it. The card also reports a small local quality suite (math 20/20, code 10/10, needles 8/8, instruction following 2/10 with reasoning_effort none) and a not-apples-to-apples overlay versus an earlier Qwen3.8 27B MLX 4-bit run on the same machine; those quality scores are not tok/s and are not listed as speed rows. mlx-lm, TensorFold and Ollama were tried the same day and did not run the official looped weights in the installed builds. Temperature 0 is not the official default (temperature 1.0).

Source: https://huggingface.co/datasets/pratikgx/inkling-2-mini-m5-pro-mac-mini-bench

Blank fields = not stated in the source. We never estimate build data; GPU street prices above are separately sourced and dated.

Embed this result

Copy into a README or blog. Links back to this build. See badge docs (CC BY 4.0 attribution). The SVG does not invent speeds β€” numbers stay on the build page.

Markdown

[![Measured on RigBase](https://getrigbase.com/badge/mac-mini-m5-pro-64gb-inkling-2-mini-iq4nl-llamacpp-pratikgx.svg)](https://getrigbase.com/builds/mac-mini-m5-pro-64gb-inkling-2-mini-iq4nl-llamacpp-pratikgx/)

HTML

<a href="https://getrigbase.com/builds/mac-mini-m5-pro-64gb-inkling-2-mini-iq4nl-llamacpp-pratikgx/" rel="noopener"><img src="https://getrigbase.com/badge/mac-mini-m5-pro-64gb-inkling-2-mini-iq4nl-llamacpp-pratikgx.svg" alt="Measured on RigBase" height="20"/></a>

Preview: Measured on RigBase Β· How to use

Rent before you buy: try this model on a cloud GPU for a few dollars first: Runpod (referral link; sign up with Google to get the new-user bonus)