Piccolo on Mac mini and MacBook Air

Research by Anamika Rawat and Tulsi Patel, supervised by Dushyant Puri, through Conestoga’s SMART Centre. This work is part of SimpleDirect®’s collaboration with the Conestoga Centre for Commercialization. I’m writing about their results; I did not conduct or independently rerun these tests.

The Mac mini generated text faster on all 24 matched prompts, averaging 45.30 tokens per second, compared with 20.29 on the MacBook Air. Across the two machines, nine responses were also marked truncated.

I’d test the Mini further before recommending a purchase. I want to see whether the faster machine delivered answers worth using.

What the two configurations produced

The researchers compared a 2023 M2 Pro Mac mini with a 2024 M3 MacBook Air with 8 GB of RAM. Both ran Vinci-Piccolo-1.0-GGUF in Q4_K_M quantization through llama.cpp, using Apple’s Metal graphics acceleration.

Piccolo is our roughly four-billion-parameter chat model, built from Qwen3.5-4B and released on June 29, 2026. The model card and download are available under Apache-2.0. “4B” describes parameter count, not a 4 GB memory requirement.

Each profile contained six different prompts. Each prompt ran once per device. The averages cover different tasks, including truncated responses; they are not repeated-trial averages. [1–4]

The first token arrived after an average 640 milliseconds on the Mini and 1,415 on the Air. The Mini was quicker on 23 of 24 prompts. A first token can belong to the model’s reasoning, so this doesn’t measure time to a useful answer.

Summed request times were 20.81 and 48.18 minutes, excluding gaps between requests. Output totals were 55,750 and 57,582 tokens. Response lengths also affected waiting time. [1, 2]

Nine responses need a closer look

All 48 requests were marked completed. Three Mini responses and six Air responses also carried truncation flags. All nine were in thinking profiles and contained exactly 4,096 tokens. The exports omit the output cap and exact finish reason, so the cause remains unconfirmed.

Other responses included 18 tokens on one Mini task and 30 on a different Air task. Without the response text or answer-quality scores, length tells us little about success. [1, 2]

What I would check before buying

Start with a task you can judge on the computer you already have. Save the full answer, the time until it becomes useful, total waiting time and any corrections you make. Repeat it in fresh contexts. Our guide to verifying AI-generated work helps define that check.

Record the exact model file, runtime build, launch command, sampling settings and output limit. Thinking and instruct used different prompts here; compare them using identical tasks.

Limits and source records

The matched inputs were 100–212 tokens; long prompts and repeat-run variation remain untested. RAM and swap readings cover the whole system. The Air already had substantial swap in its first observation, and the deck’s Mini memory label needs verification. These configurations cannot isolate chip, RAM or cooling effects, or establish battery life, wall-plug energy or throttling. [1–3]

1. Mac mini CSV (macmini_20260929_141518_all_profiles.csv): 24 observations; exclude the TOTAL PROMPT TIME summary footer.
2. MacBook Air CSV (macair_20260929_131947_Vinci-Piccolo-1.0-GGUF_Q4_K_M.csv): 24 observations, matched after whitespace normalization.
3. SimpleDirect Vinci-Piccolo Benchmarking.pptx, slides 27–37, credited to Conestoga SMART Centre.
4. Mac prompt bank (PDF) (Prompts MacAir & Mac Mini.pdf), pages 1–7.

Speed and first-token figures are per-request arithmetic means. The 2.23× speed ratio uses unrounded means; durations sum 24 request rows per device. CSVs contain no full answers or quality scores.

Read the edge-board comparison, phone test and collaboration overview.

Research credit: Anamika Rawat and Tulsi Patel, supervised by Dushyant Puri. Benchmarking through Conestoga’s SMART Centre, within SimpleDirect®’s collaboration with the Conestoga Centre for Commercialization.