Nimruz/Mizan-v2-4B
Nimruz/Mizan-v2-4B is a 4.5 billion parameter Persian-first decision model developed by Nimruz, designed to process both text and images. It excels at classification tasks, providing calibrated probabilities for allowed answers in a single forward pass without text generation or hallucinated labels. This model is particularly optimized for security camera questions, receipt analysis, and general Persian text classification, achieving high balanced accuracy on unseen images and competitive performance on various Persian text benchmarks.
Loading preview...
Mizan v2 · 4B Overview
Mizan v2 is a 4.5 billion parameter Persian-first decision model developed by Nimruz, uniquely capable of processing both text and images. Unlike generative LLMs, Mizan v2 focuses on classification, providing calibrated probabilities for predefined answers in a single forward pass. This design eliminates text generation, parsing, and hallucinated labels, making it highly reliable for specific decision-making tasks.
Key Capabilities
- Multimodal Input: Processes both text (e.g., JSON, records) and images (e.g., camera frames, receipts, product photos).
- Persian-First Optimization: Achieves strong performance on real-world Persian text data, with a mean accuracy of 82.8% across various benchmarks.
- Image Analysis: Excels at security-camera questions (e.g., person/vehicle detection, zone intrusion, counting) with 86–99% balanced accuracy, and receipt analysis (e.g., total above X?) with 99% accuracy.
- Efficient Decision Making: Returns calibrated probabilities for allowed answers in one forward pass, suitable for real-time applications.
- Flexible Deployment: Available as a text-only GGUF (2.7 GB in 4-bit) or with an additional 0.37 GB vision projector for image capabilities.
- Jev / SystemOne Compatible: Uses the same request and response body as
/v1/systemonefor easy integration.
Good For
- Security and Surveillance: Analyzing camera footage for presence detection, zone monitoring, and object identification.
- Retail and Logistics: Processing receipt photos for financial auditing and inventory management, or identifying product defects from factory photos.
- Persian Text Classification: Performing sentiment analysis, product recommendations, offensive content moderation, and intent routing on Persian language data.
- Real-time Decision Systems: Applications requiring fast, reliable classification based on multimodal inputs without the need for text generation.
- Edge Deployment: Its GGUF format and relatively small size (4.5B parameters) make it suitable for deployment on GPUs or even CPUs for less demanding tasks.