NoteTalker 2.6.1 for Mac brings two new models of our own, trained to follow on from Porites-7B. Porites2-8B comes with the app and is the default. Acropora2-4B is an optional 2.3 GB download that answers in about half the time. Both run entirely on your Mac, and you choose between them in Settings.
Both models learned from a larger and more targeted set of examples than Porites-7B, and we ran all three through the same tests, in the same app build, on the same Mac.
(Porites-7B: 146)
(Porites-7B: 1 of 4)
(Porites2-8B: 3.6 s)
(Porites2-8B: 5.06 GB)
Two names, one set of training data
NoteTalker's models are named after reef-building corals, by size. Porites is the name for our models of seven to eight billion parameters, and Acropora is the name for four billion and below. The number after the name is the generation: it goes up when the training data or method changes enough to matter. Settings shows the general-purpose model each one started from: Porites2-8B is based on Qwen3-8B, and Acropora2-4B on Qwen3-4B.
Both models learned from 4,311 recorded conversations, which make 7,933 training examples. We recorded every one of them in NoteTalker's own harness, so the model learns the app's real prompt, tools and replies. 550 of the conversations are new this round, written to teach the requests that matter most in daily use:
- a correction that changes several facts at once
- moving an event while keeping its time, place and notes
- trying again with the whole line after the app refuses a change that would drop part of it
- "keep in mind that…", which should save a fact rather than just acknowledge it
- help questions answered from the actual help pages
- questions about what you told it earlier
- plain chat that should leave your notes alone
- opening a page in the app when you ask for it by voice
Porites-7B had a fifth of the width of each layer's feed-forward block removed after training, which made it smaller and faster. We measured that trade again in each later round, on examples held back from training, and the full-width model came out ahead each time. So Porites2-8B keeps its full width: 8.2 billion parameters in a 4.6 GB file.
How we measured
Every model ran the way NoteTalker runs it: the app's prompt and settings, the same vault of example notes, and the same build of the app, on one Mac (Apple M3, 24 GB of memory). Every model ran at temperature 0, and repeat runs gave identical answers. The four test sets are:
- Everyday requests (217 single requests (136 training-style, 81 held out)).
- Everyday conversations (18 scripted conversations, 45 turns).
- Harder conversations (10 scripted conversations, 68 turns).
- General chat (30 questions that need no tools).
The conversation tests check each turn's answer and also what each note contains at the end. These test sets are newer than the ones in the Porites-7B post, so its percentages do not compare with the ones here.
The results
| Test set | Porites2-8B | Acropora2-4B | Porites-7B |
|---|---|---|---|
| Everyday requests | 158/217 · 72.8% (66.5–78.3) | 152/217 · 70.0% (63.6–75.7) | 146/217 · 67.3% (60.8–73.2) |
| Everyday conversations | 41/45 · 91.1% (79.3–96.5) | 41/45 · 91.1% (79.3–96.5) | 38/45 · 84.4% (71.2–92.3) |
| Harder conversations | 38/68 · 55.9% (44.1–67.1) | 41/68 · 60.3% (48.4–71.1) | 40/68 · 58.8% (47.0–69.7) |
| General chat | 29/30 · 96.7% (83.3–99.4) | 26/30 · 86.7% (70.3–94.7) | 27/30 · 90.0% (74.4–96.5) |
Passed / total, pass rate, and its 95% confidence interval.
What training adds
Each of our models starts from a general-purpose model, and training is what teaches it NoteTalker's job. On everyday requests, training lifts both sizes by about 20 points. On everyday conversations, Acropora2-4B passes 41 of 45 turns where the untrained Qwen3-4B passes 22.
Porites2-8B: better at the everyday
Porites2-8B passes more than Porites-7B on everyday requests (158 to 146), on everyday conversations (41 turns to 38) and on general chat (29 questions to 27). The biggest changes are in the behaviours the new data targeted:
- Remembering. Porites2-8B saves the fact on all four "remember this" cases, including "keep in mind that…". Porites-7B saves one of the four.
- Opening pages. Asking for a page by voice works on 22 of 23 cases. The model now handles it from its own training as well as through the app's shortcut.
- Saying what it did. Neither new model claimed to have saved or changed something it had not, anywhere in these tests.
Acropora2-4B: about twice as fast, at half the size
Acropora2-4B keeps its accuracy in the same range as the larger models, and it passes the most turns of the three on the harder conversations (41 of 68). It does this at half the size. We measured both new models in the same run, on the same build:
| Measure | Porites2-8B | Acropora2-4B |
|---|---|---|
| First reply, new conversation | 17.1 s | 8.4 s |
| Reply mid-conversation | 3.6 s | 1.8 s |
| Generation speed | 8.1 tokens/s | 17.8 tokens/s |
| Peak memory | 5.06 GB | 2.97 GB |
| Size on disk | 4.62 GB | 2.28 GB |
Times are medians of five rounds.
If you want quicker replies, or want to leave more memory for the other apps you have open, Acropora2-4B is the one to try.
Also in 2.6.1, for every model
Some improvements live in the app, so they help whichever model you choose:
- Asking to add something without a time now gets a question about the time, not a guess.
- If an edit would drop part of a line in your note, the app refuses it and shows the model the whole line so it can try again.
- Moving an event carries its details with it.
- A fact saved under an unusual category is filed as a general fact instead of failing.
- The in-app help now comes from the same pages as this site's help center.
Porites2-8B comes with the app. To try Acropora2-4B, open Settings, choose it and let it download once. You can switch between the two at any time. Both models run on Apple Silicon Macs.