Live demo on Hugging Face
Decisions for Bangla and Banglish, in one pass.
বাংলার জন্য তৈরি System 1 মডেল
Ask Nirnoy any set of questions about a message. It answers every one with a calibrated probability, in about a tenth of a second, without generating text.
Compatible with the System One API
How it works
Text and questions in. Probabilities out.
Questions can be yes/no, multiple choice or a 1–5 score. Pick an example.
Message
ডেলিভারি লেট হইছে, but product ta valo. refund chai na.
Nirnoy · one forward pass
Message
আমার একাউন্ট থেকে টাকা কেটে গেছে কিন্তু রিচার্জ হয়নি। এখনই দেখেন প্লিজ।
Nirnoy · one forward pass
Message
tor moto faltu lok ar dekhi nai, chup thak
Nirnoy · one forward pass
Message
Bangladesh 2nd test e 7 wicket e jitlo, Mushfiq er century 🔥
Nirnoy · one forward pass
Example outputs for illustration. Try your own text in the live demo on Hugging Face →
Any input
Text in any script, a record or a whole chat.
Send Bangla, Banglish or English, a JSON record or a list of chat turns, with one question or many. These are Nirnoy-flash's real answers. More examples →
Input · text
Nirnoy-flash · 1 question, one request
Real output · 53 ms on an L4
Why Nirnoy
Small, fast and honest about what it knows.
One pass, every question
All questions about a text are scored together in a single forward pass. Eight questions take about as long as one.
Probabilities you can trust
Calibrated with a fitted temperature, so a 0.9 means right about nine times in ten. Set thresholds and route on them.
Trained for Bangla decisions
Nirnoy starts from Google's Gemma 4 and is trained on how people really write Bangla and Banglish. On 10,000 synthetic decision scenarios it scores at least 9 points above the untrained Gemma 4 of the same size.
Traceable training data
Every training label records whether a person, a teacher model or a rule produced it, so any source can be audited or removed.
Results
First on 6 of 6 everyday tasks.
Macro-F1 (%) on 11,063 human-labelled questions, every answer option shown. Nirnoy comes in two sizes: Nirnoy-flash, small and fast, and Nirnoy 12B, the most accurate on harder tasks.
Data: MASSIVE (Bengali) (CC BY 4.0)
Data: MASSIVE (Bengali) (CC BY 4.0)
Data: BnSentMix (MIT)
Data: SIB-200 (Bengali) (CC BY-SA 4.0)
Nirnoy trained on other examples from the intent, scenario, hate and sentiment datasets; topic and emotion are new to it. By script, Nirnoy-flash scores 78.4 on Bangla and 76.4 on Banglish, against Jev's 72.6 and 61.0. The base Gemma 4 models were not run on this test with every option shown.
Benchmark
First in 9 of 9 categories against other decision models.
The everyday tasks above, scam detection on real messages and synthetic decision scenarios, against Jev, Clef, Clef-flash and the untrained Gemma 4 models Nirnoy is built on.
| Category | Nirnoy-flash | Nirnoy 12B | Jev | Clef | Clef-flash | Gemma 12B | Gemma E4B |
|---|---|---|---|---|---|---|---|
| Everyday tasks | |||||||
| Assistant commands: intent | 85.6 ★ | 82.8 | 76.6 | 79.8 | 79.5 | – | – |
| Assistant commands: scenario | 92.3 ★ | 90.8 | 84.3 | 86.9 | 88.5 | – | – |
| Hate speech in Bangla and Banglish | 84.1 ★ | 82.5 | 70.2 | 61.6 | 58.1 | – | – |
| Code-mixed sentiment | 76.2 | 77.1 ★ | 63.8 | 54.7 | 51.5 | – | – |
| News topic | 88.1 | 88.3 ★ | 88.1 | 84.3 | 79.7 | – | – |
| Emotion | 36.1 | 37.4 Tied for first: 37.4 is within 0.1 points of the best score, 37.5. The same hosted model, run again on another day, moves this much: Jev scored 37.4 on emotion on 6 October 2026 and 37.5 on emotion on 9 October 2026. So gaps this small count as a tie. | 37.5 ★ | 34.1 | 31.6 | – | – |
| Scam detection | |||||||
| SMS scam detection | 82.8 | 90.1 ★ | 88.4 | 85.3 | 83.8 | 88.6 | 83.1 |
| Scam messages, private set | 86.8 | 87.3 ★ | – | 67.3 | 59.3 | 73.0 | 82.6 |
| Decision scenarios, synthetic (accuracy) | |||||||
| Decision scenarios | 77.2 | 80.1 ★ | 75.7 | 74.6 | 67.7 | 70.8 | 67.8 |
Macro-F1 (%); accuracy for the decision scenarios. ★ marks the best model in each row; scores within 0.1 points of the best count as a tie and share it, because the same hosted model moved that much between two runs (Jev scored 37.4 on emotion on 6 October 2026 and 37.5 on emotion on 9 October 2026). Every model saw the same text, question and answer options. Jev, Clef and Clef-flash answered through their public APIs (OpenRouter, October 2026), except on the private scam messages, which never leave our machines: there Clef and Clef-flash ran from their released weights on our GPUs and Jev was not run. The Gemma 4 rows are the untrained models Nirnoy is built on (not run on the everyday tasks with every option shown). Nirnoy never trained on the scam sets, and the decision scenarios have synthetic answers. Every category, with an example →
API · coming soon
One request per message.
The hosted API is not open yet. It speaks the System One wire format, so it will drop in behind any client that already uses it. Need it sooner? Ask for early access.
curl https://[host]/v1/systemone \
-H "Content-Type: application/json" \
-d '{
"state": "ডেলিভারি লেট হইছে, but product ta valo.",
"questions": {
"sentiment": {
"type": "choice",
"instructions": "Overall sentiment?",
"criteria": ["positive", "negative", "mixed", "neutral"]
},
"refund": {
"type": "noul",
"instructions": "Does the customer ask for a refund?"
}
}
}'{
"model": "nirnoy-flash",
"answers": {
"sentiment": {
"type": "choice",
"choice": "mixed",
"probabilities": {
"positive": 0.21, "negative": 0.14,
"mixed": 0.62, "neutral": 0.03
}
},
"refund": { "type": "noul", "noul": false }
},
"usage": { "input_tokens": 212, "output_tokens": 0 }
}Coming soon
The API is coming soon.
Today you can try Nirnoy-flash and Nirnoy in the live demo on Hugging Face. The hosted API is not open yet. If you need it for your product, tell us about it: where it suits both sides, we can give you early access.
Try it on Hugging Face