Hermes Agent · Mac mini
Hermes Agent on a Mac mini: what 16 GB actually does
Short answer: On one narrow job, a 16 GB Mac mini M4 got 15 of 15 unseen quote requests right, averaging 1.1 seconds each (6 Oct 2026).
On a new trade it had never seen, an electrical distributor (14 Oct, Dell test box), the default Qwen3 8B setup scored 30 of 40 and 2 of 15 on messy requests, with 8 wrong prices in 55, because per-foot units, price levels and trade shorthand are not handled yet; that is why we score your own list first (the electrical result).
A 16 GB Mac mini will start Hermes Agent; most write-ups stop there. Whether it is worth putting in an office depends on three numbers they skip: how often it is right, how long each answer takes, and where it runs out of room. We measured all three on our demo unit with Qwen3 8B and say plainly what we have not measured. Want the same numbers on your own requests? Tell us in the form below; a person on our team replies.
Updated 14 October 2026 · built on Hermes Agent, not affiliated with Nous Research
Staffbox is built on Hermes Agent (Nous Research, MIT license). Staffbox is not affiliated with Nous Research.
On 16 GB, Qwen3 8B got 15 of 15 unseen requests right (6 Oct) and no total wrong in 475 answers
Every run below used Qwen3 8B on the 16 GB M4 with a brain and a quote action, one request at a time with the model warm. It is not a general-purpose assistant, and we have not tested it as one. Companies are fictional: Fieldstone IT is an IT provider, Peachtree Cabinet Works is a cabinet shop.
| Date | Test set | Score | Average / 90th percentile |
|---|---|---|---|
| 6 Oct 2026 | Messy set 2, unseen | 15/15 | 1.1 s / 1.7 s |
| 6 Oct 2026 | Messy set 1, seen while fixing | 19/20 | 1.2 s / 2.3 s |
| 6 Oct 2026 | Fieldstone IT | 40/40 | 2.3 s / 5.3 s |
| 6 Oct 2026 | Peachtree Cabinet Works | 39/40 | 2.5 s / 4.4 s |
| 6 Oct 2026 | Generated regression set, Fieldstone | 201/204 | 1.6 s / 2.1 s |
| 6 Oct 2026 | Generated regression set, Peachtree | 150/156 | 3.2 s / 3.4 s |
Of the 11 misses in the 6 October run, 10 declined and asked the customer, and 1 was a routing miss. Declining counts as a miss on our scorecard, which is deliberate: a refused request is work for a person. The generated regression set is requests written by a local model on a separate machine, with answers checked by code; it guards against regressions and is not the accuracy proof. Fifteen messy requests is a small sample.
Average answers took 1.1 to 3.2 seconds, and nine in ten arrived within 5.3 seconds (6 Oct)
Speed is the second thing an office notices. The chart puts each test set's average next to its 90th percentile, the time nine answers in ten came in under. The gap is widest on the clean Fieldstone set: 2.3 seconds on average, 5.3 at the 90th percentile (6 Oct 2026).
These timings assume a warm model. We have not measured cold starts, which is why the unit keeps the model loaded all day; the settings are on the Ollama page.
The limits are one request at a time and no room for a bigger model
- Concurrency. One request at a time. Concurrent requests queue. We have not measured several staff using it at once.
- Bigger models. A 14B model timed out on 16 GB (demo-unit notes, 1 Oct 2026), so we do not use it in demos.
- Changing context size. Changing num_ctx costs about 100 seconds on 16 GB (demo-unit notes, 1 Oct 2026). We keep one context size per run. See the 64K context page.
- Hangs. Ollama can hang when the disk is busy or models unload between steps. The fix and the settings that avoid it are on the Ollama page.
- Without a brain. The same Qwen3 8B scored 1/40 on Fieldstone IT without the brain, on a different machine (Dell RTX 3090, 30 September 2026). The model is not the product; the notes and the quote tool are.
If you try this on your own Mac mini, these steps steer around most of those limits:
- Set the Ollama keep-alive to 24 hours so the model stays loaded.
- Fix one context size and do not change it between runs.
- Bind Ollama to 127.0.0.1.
- Write 30 real requests and a held-out 10 before you tune anything, and grade the 10 blind.
- Score without the notes first, so you know how much the notes are worth.
If you would rather we ran the 30 for you, that is the free test. More on local hardware at local AI on a Mac mini.
16 GB fits the 8B model; the non-Qwen option and customer pilots need 32 GB
| Model class | Memory | What we know |
|---|---|---|
| Qwen3 8B | 16 GB works | Scores above. Speed measured on the M4. |
| gpt-oss-20b (non-Qwen option) | Needs 32 GB | Scored 39/40 Fieldstone, 40/40 Peachtree, 15/15 messy on a Dell RTX 3090 (PR #7, 2 Oct 2026). Speed on Apple hardware is not measured yet. |
| Qwen3 27B class | Mac Studio planned | 40/40 and 39/40 on the Dell with brain and quote action (30 Sep 2026). Mac Studio speed is not measured. |
All Mac mini scores here are on a 16 GB unit. Customer pilots may run on 32 GB (the install guide: "32 GB memory recommended; 16 GB runs the 8B model"); nothing is scored on 32 GB yet, so a pilot unit is measured on the same tests before day 0.
Running cost is small either way. Apple rates the Mac mini (M4) at 4 W idle and 65 W maximum (Apple support document 103253). We have not measured our own unit's draw during quoting, so we quote Apple's rating only.
Related: Hermes Agent with Ollama (the settings the demo unit uses) · Local models for Hermes Agent (scores for Qwen3 8B, a 27B model and gpt-oss-20b) · Hermes Agent and 64K context (the 65,536 setting every score was measured at) · Install Hermes Agent on a Mac mini (the installer, step by step) · Hermes Agent FAQ (short answers to common questions).
Back to the Hermes Agent hub. See also models, the free AI quote test.
Questions
What Mac mini do I need for Hermes Agent?
For our quoting task, a 16 GB Mac mini M4 running Qwen3 8B scored 40/40 and 39/40 on two fictional companies (6 Oct 2026). The non-Qwen option, gpt-oss-20b, needs 32 GB. We have not tested other uses.
Is a Mac mini enough for several users at once?
Not measured. Our tests run one request at a time, and concurrent requests queue.
How fast is Hermes Agent on a Mac mini M4?
On our quoting task with Qwen3 8B the average was 1.1 to 3.2 seconds per answer in the 6 October runs, with the model warm. The 90th percentile was 1.7 to 5.3 seconds. We have not measured cold starts. A 14B model timed out on the same 16 GB machine.
Sources
- Hermes Agent guide: local LLM on Mac, Nous Research, seen 13 Oct 2026
- Apple: Mac mini (M4) power consumption
- github.com/Staffbox-ai/staffbox (Staffbox code and evals/results, MIT), evals/results 2026-10-02 to 2026-10-06
- huggingface.co/spaces/staffbox/scorecard (every graded answer)
For the business case, read why a small business runs local AI on a Mac mini. To set one up, follow installing Hermes Agent on a Mac mini with the Staffbox installer.
Or call us: 404 493 6248