munzurul/speaklar_gemma-3-1b-it

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:1BQuant:BF16Context Size:32kPublished:Sep 11, 2026Architecture:Transformer Featherless Exclusive Cold

The munzurul/speaklar_gemma-3-1b-it is a 1 billion parameter Gemma 3 instruction-tuned model developed by munzurul. It is specifically fine-tuned for grounded Bengali voice-bot conversations, excelling at answering from supplied evidence, handling unavailable information, and managing voice-agent style constraints. This model is optimized for tasks requiring detailed information gathering for orders and appointments, and safely managing sensitive requests.

Loading preview...

Overview

The munzurul/speaklar_gemma-3-1b-it is a 1 billion parameter Gemma 3 instruction-tuned model, specifically fine-tuned for grounded Bengali voice-bot conversations. It is designed to operate within a knowledge-base context, providing answers based on supplied evidence and indicating when information is unavailable.

Key Capabilities

  • Contextual Grounding: Answers questions strictly from provided knowledge-base context.
  • Bengali Voice-Agent Style: Adheres to specific stylistic constraints for Bengali voice agents.
  • Information Gathering: Efficiently gathers essential details for orders and appointments.
  • Safe Handoff: Capable of safely handing off sensitive or unsupported requests.
  • Multilingual Input: Trained on Bengali, English, and Banglish inputs, with Bengali-only assistant outputs.

Training Details

The model was fine-tuned using QLoRA on a 15,000-example Bengali voice-bot behavior dataset. This dataset covers a range of scenarios including evidence-grounded answers, abstention, calculations, order/appointment intake, payment safety, complaint intake, and human handoff.

Good For

  • Developing Bengali voice-bots that require grounded responses.
  • Applications needing to extract specific information for transactional purposes (e.g., booking, ordering).
  • Systems where safe and controlled responses are critical, especially for sensitive user interactions.