DeepSeek R1 32B vs M4 Mac Mini Pro 24GB

Product reviewed

• DeepSeek R1 (32B) run locally on a Mac mini Pro (M4 Pro, 24GB RAM) (base model).

System / setup details mentioned

• Mac mini Pro: M4 Pro, 24GB RAM
• Pricing reference: ~£1300 UK / ~$1400 US
• DeepSeek model sizes discussed:
  • 70B parameter version
    • Technically ran, but very slow
    • Attempt result: ~27-minute response time to a “hello”
    • Model size: ~43GB download
    • Practical use: limited, unless you “have all day”
  • 32B parameter distilled model (used for all subsequent prompts)
    • “Distilled” explanation: DeepSeek trains smaller models via reasoning from a larger system (described as training 671B parameters)

Test method & prompts

• Ran DeepSeek R1 32B using Terminal
• Monitored performance with:
  • Activity Monitor (memory usage graph)
  • GPU usage (described as being “heavily hammered”)
• Verbose mode enabled to capture timing/metrics
• Three prompt examples:
  1. A complex question aimed at “making it simple”
  2. A coding example / prompt example generation
  3. Requesting lyrics for The Beatles – “A Hard Day’s Night” (specific, well-known pop song)

Key results (timings)

1. Prompt 1: ~1 minute 40 seconds
   • Completion described as reasonable
2. Prompt 2: ~3 minutes 56 seconds
   • Behavior:
     • Used all physical memory
     • Reached “Amber” (but not “Red” for any significant time)
     • GPU was heavily hammered in chunks
   • Conclusion: 32B appears near the comfortable “top end/sweet spot” for this Mac mini M4 Pro (24GB)
3. Prompt 3: ~3 minutes 25 seconds
   • Issue: model hallucinated and produced incorrect lyrics
   • It did not provide “A Hard Day’s Night”; instead responded with something like “a day in the life” (“gobble deg… nonsense”)

Main features / takeaways about the product experience

• Usable locally with 32B on the M4 Pro / 24GB Mac mini
• Performance sweet spot: the reviewer implies 32B is “there or thereabouts” for what the machine can comfortably handle
• Memory pressure is real:
  • Full RAM usage during longer tasks
  • Stress increases over time, reaching Amber but not sustained Red
• No internet search in this setup
  • Likely contributes to hallucinations when asked for exact song lyrics

Pros (as implied by the review)

• Good enough responsiveness for practical use at 32B
  • ~1:40 for the complex simplification question
• Runs stably without sustained critical memory throttling
  • Hits Amber, but avoids Red for long periods
• Coding-related behavior
  • Reviewer notes smaller DeepSeek models tend to produce code examples when prompted; the 32B run was expected to show similar behavior (and was interesting to compare)

Cons (as implied by the review)

• 70B is impractical on this hardware
  • ~27 minutes even for “hello”
  • Large download (~43GB) and poor time-to-answer
• Lyrics task reliability is poor
  • Hallucinations and wrong song selection
  • Likely worsened by no browsing/search capability

Comparisons made

• DeepSeek 70B vs 32B on the same Mac mini
  • 70B: works but massively slower (~27 minutes) → not practical
  • 32B: reasonable performance (minutes) with manageable stability → practical sweet spot
• Smaller variants vs 32B (coding prompt behavior)
  • Smaller models (mentioned: 1.5B and 7B) “struggle” generating prompt examples and tend to output code examples instead
  • The 32B test was intended to check if this improves
• Within-setup constraint
  • Even if the model may know lyrics exist in training data, it still struggles to be correct without external verification (since it can’t search the internet)

Overall recommendation / verdict

• Recommended for local experimentation and general reasoning/coding prompts at 32B on a Mac mini M4 Pro 24GB
• Not recommended to run DeepSeek 70B on this machine due to:
  • Extreme latency (~27 minutes)
  • Very large download
• Not reliable for exact copyrighted lyric reproduction in this setup; expect hallucinations, especially since it can’t search online

Unique points mentioned (consolidated list)

• Mac mini Pro (M4 Pro 24GB) used for local DeepSeek R1 testing
• 70B model:
  • ~27-minute response to “hello”
  • ~43GB download
  • Technically runs, but impractical
• 32B distilled model selected as the practical choice
• Distillation described as training smaller models from larger reasoning capacity (reviewer mentions 671B parameters)
• Three prompt types tested: complex simplification, coding/examples, and Beatles lyrics
• Timing results: 1:40, 3:56, 3:25
• Performance/stability:
  • Memory reached Amber and used all RAM, but avoided sustained Red
  • GPU heavily used during portions of the runs
  • 32B found near the sweet spot/top end for this hardware
• Lyrics task:
  • Model hallucinated
  • Returned content inconsistent with requested song (A Hard Day’s Night), leaning toward “a day in the life”
• Lack of internet access limits correctness

Speakers / voices

• Single speaker (no separate contributors indicated in the subtitles).