ZANA Cognitive Engine: runtime models
The small trained components that the ZANA Cognitive Engine desktop app ships with. They are tiny, task-specific models for an on-device assistant. They are not a general chat or instruction-following language model. Open-ended answers in the ZANA app come from a separate local model (for example one served by Ollama), if you run one.
| File | What it is | Size |
|---|---|---|
intent_v3_C_s0.safetensors (+ .json) |
Intent classifier used by the default composite front end: decides what the user is asking for (remember, recall, list files, create a task, calculate, search documents, …). 449,555 parameters, d_model 128, seed 0. |
1.8 MB |
calib_b_frozen.json |
Frozen per-intent calibration thresholds used together with the classifier. | 2.5 KB |
core_lm_v1.safetensors (+ .json) |
ZANA's recurrent cognitive core as a byte-level language model (SSM + attention, d_model 64, 132,416 parameters), trained on English, Arabic and Sorani Kurdish text. |
0.5 MB |
Evaluation (from the training records in the .json files)
- Intent classifier: best accuracy 0.804 on the project's dev split (
best_dev3_acc). It was never trained on the project's evaluation sets, MASSIVE dev/test, or the held-out Common Voice splits listed inintent_v3_C_s0.json(never_trained_on). This is a development-split number from training, not a published benchmark. - Core language model: held-out bits per token after 3,000 steps: English 2.23, Sorani Kurdish 1.61, Arabic 1.62. The held-out text was never trained on.
Training data
- Intent classifier: 600 templates plus 90 natural seed utterances and a 797-example training split (each repeated 3 times), as recorded in
intent_v3_C_s0.json. - Core language model: Common Voice sentence corpora for English, Arabic and Sorani Kurdish (CC0) and the MASSIVE English and Arabic training splits (CC BY 4.0, © Amazon; credited here as the licence requires). See
DATASETS.mdin the GitHub repository.
Use
from safetensors.torch import load_file
weights = load_file("intent_v3_C_s0.safetensors")
The model architectures and the code that loads these files are in the GitHub repository (src/zana/). The easiest way to use them is the ZANA desktop app, which bundles them.
Limits
- The intent classifier covers ZANA's own set of assistant actions and was tuned for English, Arabic and Sorani Kurdish short requests. It will misclassify requests outside that set.
- The core language model is very small (132 K parameters). It is a research component of the engine, not something to generate text with.
- No claims are made about performance outside the numbers above.
App downloads (preview)
The ZANA desktop app, with these models bundled, is in apps/:
| File | For | SHA-256 |
|---|---|---|
apps/ZANA.dmg (560 MB) |
macOS, Apple silicon | 0fd2c47074240f176e0ffa8827b593aa95a8ebbeb11c7b341f7418fc25b102c6 |
apps/ZANA-Windows-x64.zip (484 MB) |
Windows 64-bit | 5a52ca6487a0c20315ca45dd9504bc50db71ee9930ebda8944fa8ff7e5e291d2 |
- Not signed by a developer certificate. On macOS, right-click the app and choose Open the first time. On Windows, SmartScreen may warn: choose More info, then Run anyway.
- Preview limits: no voice playback or cancel button yet; the Windows app was only start-up tested on a hosted Windows machine, not opened in a real window; microphone input and a real local language model were not tested.
- For full answers to open questions, install Ollama and run
ollama pull qwen2.5:7b-instruct; ZANA finds it automatically. - Source, release notes and build scripts: https://github.com/danamohammad/ZANA-Cognitive-Engine (private; ask the owner for access).
License
Apache 2.0, the same as the code. The datasets above keep their own licences.