On a 16 GB Mac mini with an Apple M4, the Grux Local 2B writes about 48.0 tokens a second, the 4B about 24.6 and the 9B about 18.3, measured on 17 November 2026 through Ollama 0.40.8. A token is roughly three quarters of a word, so the 4B writes about 18.5 words a second, where a person reads about 4. With the model already loaded, the first word starts in well under a second at every size, and a 200 token reply, about 150 words, finishes in 4.2 seconds on the 2B, 8.3 on the 4B and 11.1 on the 9B. Speed falls as the model grows: the 2B is 2.6 times faster than the 9B. These are one Mac, one runner and short replies. A longer prompt, a longer context or a busier Mac reads slower, and a 24 GB Mac, the size the 9B is made for, was not measured.
Each row is the median of 10 timed replies from a model already in memory, on a 16 GB Mac mini with an Apple M4, macOS 26.3.1. Speed is the tokens Ollama generated divided by the time Ollama spent generating them, so it leaves out loading and reading the prompt. Across the 10 replies the speed stayed within 47.7 to 48.4 on the 2B, 22.6 to 25.8 on the 4B and 17.9 to 18.4 on the 9B, and the median time to the first word was 0.05, 0.09 and 0.12 seconds.
| Size | Tokens a second | 200 tokens |
|---|---|---|
| 2B, for 8 GB | 48.0 | 4.2 s |
| 4B, for 16 GB | 24.6 | 8.3 s |
| 9B, for 24 GB | 18.3 | 11.1 s |
Tokens are not words, so this table converts at three quarters of a word a token, a rule of thumb for English, and compares with a reader at 240 words a minute, which is 4 words a second.
| Size | Writes | Against a reader at 240 words a minute |
|---|---|---|
| 2B | 36.0 words a second | 9.0 times |
| 4B | 18.5 words a second | 4.6 times |
| 9B | 13.7 words a second | 3.4 times |
gruxai.com/grux/grux-local:2b, :4b and :9b, pulled from gruxai.com.ollama run gruxai.com/grux/grux-local:4b --verbose prints the eval rate, in tokens a second, after every reply.| Claim | Source |
|---|---|
| Speed is the tokens generated over the time spent generating them | Ollama API reference, the eval_count and eval_duration fields |
| The three tags and what each is made for | Grux Local |
| Which Ollama models fit which memory | Ollama models and which LLM can my Mac run |