Nimruz/Mizan-v3-3B
Text Classification • 3B • Updated • 227 • 19
Mizan: Persian decision models
Persian-first decision models by Nimruz: typed questions in, calibrated answers out, one forward pass. Start with Mizan v3.
Note START HERE. Text & JSON. Follows the policies and scoring rubrics you put in the record. 82.5% on real Persian data (level with 9B–27B models), 91.1 on Tasmim-Bench, best calibration. The card has a 5-minute guide.
Nimruz/Mizan-v3-3B-GGUF
Text Classification • 3B • Updated • 770 • 8Note Mizan v3 for llama.cpp: 1.9 GB (Q4) or 3.1 GB (Q8). Runs on an 8 GB GPU or a laptop CPU, ~55 ms per request on an L4.
Nimruz/Mizan-v2-4B
Image-Text-to-Text • 5B • Updated • 95 • 2Note Use this when your records include IMAGES: camera frames, receipts, invoices, factory and product photos. Text plus vision.
Nimruz/Mizan-v2-4B-GGUF
Image-Text-to-Text • 4B • Updated • 545 • 2Note Mizan v2 for llama.cpp: 2.7 GB model + 0.37 GB vision projector (only needed for images).
Mizan Demo
⚖Live demo of Mizan, a Persian decision model
Note Try Mizan v2 in the browser with Persian examples: text, camera frames, factory photos and receipts.