Banglish Companion — MLX
8-bit MLX build of banglish-companion, for Macs. What the model is and why it exists are on that page. This one is about running it.
1.9 GB. No install beyond mlx-lm, no account, no network calls once it is on
disk.
There is a newer model. banglish-companion-3b is built on a 3B base and answers shorter and closer to the question. On a Mac, banglish-companion-3b-mlx runs it at 3.6 GB. These files stay published and working; start there if you are choosing today.
Banglish e
ki eta? banglish-companion er MLX version — Mac er jonno. 1.9 GB.
ki lagbe? pip install mlx-lm. Apple silicon Mac lagbe. Internet lagbe na,
account lagbe na, taka lagbe na. Tumi ja bolba, tomar computer er baire jabe na.
bipod hole? kichu message er uttor model dey na — companion.py age
oigulo dekhe ney, tomar kichu korte hobe na.
Use it
pip install mlx-lm huggingface_hub
hf download sifat-febo/bangla-banglish-companion-mlx --local-dir banglish
python banglish/companion.py
That starts a chat in your terminal. It first answers aj mood bhalo na
("I'm not in a good mood today"), then whatever you type. In your own code,
from companion import reply is the only import you need. It reads the
message first with the small ontor classifier that ships here. If someone
writes that they are in danger, it replies with fixed text and does not ask
the model. Otherwise it answers with the settings this model was tested with.
Apple silicon only — MLX does not run on Intel Macs or on Linux. For those, use the GGUF builds.
What answers you, and when
If someone writes that they are in danger, the reply is fixed text from
ontor/nirapotta.py, not the model, and reply() checks that first on every
turn. It is plain Python, needs nothing installed, and runs in microseconds.
Why 8-bit and not 4-bit
We converted both and checked them against the weights they came from. 8-bit matched its source closely. 4-bit did not — it lost noticeably more on emotional conversations and started producing confidently wrong answers, so it is not published. The 4-bit GGUF holds up better than 4-bit MLX does; if you need something smaller than this file, take that one.
Limits
- Banglish only. Romanized Bengali. Not Bengali script, not English.
- May hallucinate. It is a 1.7B model. It will state wrong things with confidence.
- Not a professional advisor. Not a therapist, doctor, or lawyer.
- Short exchanges. It degrades over long conversations — it repeats itself and loses the thread.
- It deflects rather than refuses. On a harmful request it tends to change the subject instead of declining outright. That is not the same thing as a refusal. Do not put this anywhere a refusal matters.
- Some answers are not the model's. If someone writes that they are in danger, the reply is fixed text from a file in this repo. The model is not asked.
AI Disclosure
The author designed the system, chose the base models, ran all training and measurement on the author's own machine, read the models' conversations, and chose what to publish and what to withhold. Claude Code (Anthropic) was used as a coding and writing tool under that direction.
License
Apache 2.0
Cite the technical report, which describes how the models were built, trained and measured: https://doi.org/10.5281/zenodo.22371896
@misc{febo2026banglishcompanion,
author = {Febo, Sifat},
title = {Banglish Companion: Small Chat Models for Romanized Bengali
— Technical Report},
year = {2026},
month = sep,
publisher = {Zenodo},
version = {1.0},
doi = {10.5281/zenodo.22371896},
url = {https://doi.org/10.5281/zenodo.22371896},
note = {Repository: sifat-febo/bangla-banglish-companion-mlx}
}