dheeyantra/dhee-pravega

VISIONConcurrent Unit Cost:1Model Size:5.1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 26, 2026License:gemmaArchitecture:Transformer0.0K Featherless Exclusive Cold

Dhee-Pravega is a 5.1 billion parameter agentic tool-calling model developed by DheeYantra Research Labs, built upon the Gemma-4 architecture. It is specifically optimized for low-latency tool-calling and multilingual interaction across 13 Indian languages plus English, utilizing Gemma-4's native tool-calling format. This model excels at generating correct tool calls and replying in the user's native language, making it ideal for in-language voice and chat agents.

Loading preview...

Dhee-Pravega: Agentic Tool-Calling for Indian Languages

Dhee-Pravega is a compact, low-latency agentic tool-calling model developed by DheeYantra Research Labs, designed to facilitate in-language voice and chat agents across 13 Indian languages. Built on the Gemma-4 architecture, it leverages Gemma-4's native tool-calling format for seamless integration with standard tokenizer.apply_chat_template workflows.

Key Capabilities & Features

  • Multilingual Tool-Calling: Supports 13 Indian languages (Assamese, Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Sanskrit, Sindhi, Tamil, Telugu) plus English, accurately picking functions and replying in the user's language.
  • Ultra-Low Latency: Achieves sub-50 ms to first token on Blackwell GPUs, with verified ~47 ms on an RTX 4090 for tool-calling prompts, enabling near-instant responses.
  • Language Consistency: Maintains replies in the user's native language, even when system prompts and tool schemas are provided in English.
  • Gemma-4 Base: Inherits the robust foundation of google/gemma-4-E2B-it and its associated terms of use.

Ideal Use Cases

  • In-Language Voice Bots (IVR): Rapidly process user requests and execute tool calls in their native language.
  • Chat Assistants: Develop responsive and context-aware chat agents for diverse Indian language users.
  • On-Device Agents: Suitable for applications requiring efficient, localized agentic capabilities where English-only models fall short.

Limitations

  • Best suited for short, tool-oriented exchanges.
  • Tool-calling fluency is strongest for schemas resembling the Glaive-style training distribution.
  • Subject to the Gemma base model's terms and prohibited-use policy. This non-commercial release requires a separate license for commercial use.

Finetunes
1 models