OpenClaw / Ollama — Hardware Field Report Research conducted June 2026  ·  Matched to prior LLM inference series

Mac mini 2020–2024 — OpenClaw Performance Projection

Mac mini 2020–2024: five chip generations, ten configurations

The question isn't whether OpenClaw runs — it's whether it runs well enough for what your colleague actually wants to do with it.

Chip generations
5
Configurations
10
Operating modes
2
Model size range
7–32B
SECTION A

OpenClaw runs in two modes

Cloud API Mode Hardware-independent
How it works OpenClaw routes queries to Claude, ChatGPT, or Gemini over the internet. The Node.js gateway process itself is lightweight (~200–400 MB RAM, minimal CPU). Local hardware is essentially irrelevant.
Takeaway Any Mac mini on this list handles cloud mode identically. Step time ≈ 30–90 sec for a 10-step task — API latency, not hardware.IDENTICAL
Local Ollama Mode Hardware-bound
How it works OpenClaw calls a local Ollama instance. Everything below applies to this mode only. Performance is entirely determined by: (1) how much RAM you have — controls which models fit, and (2) memory bandwidth — controls how fast tokens generate.
Takeaway Node.js adds ~200 MB overhead on top of the Ollama model. Plan RAM budget accordingly.PLAN RAM
SECTION B

Memory bandwidth

All Mac mini generations + M1 Max reference

ChipMemory bandwidthReleased
M168.25 GB/s2020
M2100 GB/s2023
M2 Pro200 GB/s2023
M4120 GB/s2024
M4 Pro273 GB/s2024
M1 Max (reference)400 GB/sPrior report — not a Mac mini
Note

M1 Max shown as reference from prior report (not a Mac mini). Token generation scales linearly with bandwidth for memory-bound models.

SECTION C

The key finding

The prior report found qwen3:32B timed out on bank sessions because it generated only 10–13 tok/s on an M1 Max. That problem disappears with smaller models. A 7B model generates 18–52 tok/s depending on the chip — fast enough that a 10-step browser task completes in 3–5 minutes on every Mac mini here, well inside a 10–15 minute bank session timeout. The constraint shifts from speed to capability: can the model you can fit in your RAM actually navigate complex authenticated flows reliably?

SECTION D

Configuration detail

Local Ollama mode · best-fit model per RAM tier

Mac mini M1  ·  68.25 GB/s  ·  NOV 2020

8 GB~5.5 GB for model
Model tiers7B ✓14B32B
qwen3:7b Q4 recommended
Speed18 tok/s (7B)
10-step time4.3 min / 10 steps
VerdictCloud APIGOOD · 7B local (marginal quality)MARGINAL · 14B won't fitNO FIT · Browser auto (limited)LIMITED
16 GB~13.5 GB for model
Model tiers7B ✓14B ~32B
14B fits but tight; monitor swap
Speed10 tok/s (14B)
10-step time5.8 min / 10 steps
VerdictCloud APIGOOD · 7B localGOOD · 14B (tight fit, slow)TIGHT · Browser auto (usable)USABLE

Mac mini M2  ·  100 GB/s  ·  JAN 2023

8 GB~5.5 GB for model
Model tiers7B ✓14B32B
faster than M1 8GB — same model ceiling
Speed25 tok/s (7B)
10-step time3.8 min / 10 steps
VerdictCloud APIGOOD · 7B local (marginal quality)MARGINAL · 14B won't fitNO FIT · Browser auto (limited)LIMITED
16 GB~13.5 GB for model
Model tiers7B ✓14B ~32B
14B tight; 46% faster than M1 at 14B
Speed14 tok/s (14B)
10-step time4.9 min / 10 steps
VerdictCloud APIGOOD · 7B localGOOD · 14B (tight fit)TIGHT · Browser auto (decent)DECENT

Mac mini M2 Pro  ·  200 GB/s  ·  JAN 2023

16 GB~13.5 GB for model
Model tiers7B ✓14B ~32B
14B tight but faster here than M2 16GB
Speed22 tok/s (14B)
10-step time4.0 min / 10 steps
VerdictCloud APIGOOD · 7B local (fast)FAST · 14B (fast but tight)TIGHT · Browser auto (good)GOOD
32 GB~29.5 GB for model
Model tiers7B ✓14B ✓32B
14B comfortable; 32B still won't fit
Speed22 tok/s (14B)
10-step time4.0 min / 10 steps
VerdictCloud APIGOOD · 7B local (fast)FAST · 14B (fast, comfortable)GOOD · Browser auto (very good)VERY GOOD

Mac mini M4  ·  120 GB/s  ·  NOV 2024

16 GB~13.5 GB for model
Model tiers7B ✓14B ~32B
M4 base is now the cheapest new Mac mini
Speed16 tok/s (14B)
10-step time4.6 min / 10 steps
VerdictCloud APIGOOD · 7B local (fast)FAST · 14B (tight fit)TIGHT · Browser auto (good)GOOD
24 GB~21.5 GB for model
Model tiers7B ✓14B ✓32B
14B comfortable; sweet spot for M4
Speed18 tok/s (14B)
10-step time4.4 min / 10 steps
VerdictCloud APIGOOD · 7B local (fast)FAST · 14B (comfortable)GOOD · Browser auto (very good)VERY GOOD

Mac mini M4 Pro  ·  273 GB/s  ·  NOV 2024

24 GB~21.5 GB for model
Model tiers7B ✓14B ✓32B
fastest 14B of all configs on this list
Speed30 tok/s (14B)
10-step time3.6 min / 10 steps
VerdictCloud APIGOOD · 7B local (very fast)VERY FAST · 14B (comfortable, fast)GOOD · Browser auto (excellent)EXCELLENT
48 GB~45.5 GB for model
Model tiers7B ✓14B ✓32B ✓
only config on this list that runs 32B
Speed8 tok/s (32B)
10-step time6.8 min / 10 steps
VerdictCloud APIGOOD · 7B / 14B (very fast)VERY FAST · 32B (fits, slower)SLOWER · Browser auto (excellent)EXCELLENT
SECTION E

All 10 configurations at a glance

Config Bandwidth Best local model Tok/s 10-step time Cloud 7B 14B Browser auto
M1 · 68.25 GB/s · Nov 2020
M1 / 8 GB68 GB/s7B Q4184.3 min ●● marginal● won't fit●
M1 / 16 GB68 GB/s14B Q4105.8 min ●●● tight●
M2 · 100 GB/s · Jan 2023
M2 / 8 GB100 GB/s7B Q4253.8 min ●● marginal● won't fit●
M2 / 16 GB100 GB/s14B Q4144.9 min ●●● tight●
M2 Pro · 200 GB/s · Jan 2023
M2 Pro / 16 GB200 GB/s14B Q4224.0 min ●●● tight, fast●
M2 Pro / 32 GB200 GB/s14B Q4224.0 min ●●●●
M4 · 120 GB/s · Nov 2024
M4 / 16 GB120 GB/s14B Q4164.6 min ●●● tight●
M4 / 24 GB120 GB/s14B Q4184.4 min ●●●●
M4 Pro · 273 GB/s · Nov 2024
M4 Pro / 24 GB273 GB/s14B Q4303.6 min ●●●●
M4 Pro / 48 GB273 GB/s32B Q486.8 min ●●●●

Legend: green ● = good · amber ● = marginal/tight · red ● = won't fit / not viable

SECTION F

Recommendation — match to your colleague's machine

Every Mac mini on this list can run OpenClaw. What varies is the quality of the local model it can run.

Cloud API mode — Any configuration

  • Hardware is irrelevant. Performance is identical across all ten configs. If your colleague just wants OpenClaw working with Claude or ChatGPT in their messaging apps, any Mac mini suffices.

Local model — 8 GB configs (M1 8 GB · M2 8 GB)

  • 7B only. Fast enough for simple flows. Not reliable for complex authenticated bank navigation — the model will hallucinate or lose context mid-flow. Cloud mode is the right answer here.

Local model — 14B sweet spot (M2 Pro 32 GB · M4 24 GB · M4 Pro 24 GB)

  • 14B fits comfortably, generates in 18–30 tok/s, and completes a 10-step bank task in 3.6–4.4 min. This is the practical minimum for reliable autonomous browser flows.

Local 7B, passable for simple flows (M1 16 GB · M2 16 GB · M2 Pro 16 GB · M4 16 GB)

  • 14B is tight but runs. Use qwen3:14b with thinking disabled. All complete well inside the bank timeout window. Worth testing before investing in more RAM.

For your colleague's specific question — setup time is not the risk. OpenClaw installs in minutes on macOS; Ollama pulls a model in one command. The risk is discovering mid-task that their Mac mini doesn't have enough RAM to run a model capable of what they're trying to automate.

Start with cloud API mode. It's zero setup overhead and lets your colleague learn OpenClaw's behavior before betting on a local model. Once they know what they need it to do, they'll know whether a 7B model is sufficient or whether they need 16 GB+ to run a 14B.

The one configuration to avoid for serious local use: 8 GB on any chip. macOS itself consumes 2–3 GB, leaving under 6 GB for a model — enough for a 7B but not for the quality level needed to navigate an authenticated multi-step bank session reliably.

Sources

Apple Support spec pages llama.cpp GitHub discussion #4167 localaimaster.com openclawai.io hardware guides Prior report series (M5 Max inference field report, Jun 2026)