ritiksuman/rudra

Rudra v1: A 4.5B Coding Agent & Reasoning LLM

Rudra is a 4.5 billion parameter language model developed by Vyomma Intelligence, with Ritik Suman as the brain behind its creation. It is a fine-tune of unsloth/Qwen3.5-4B using QLoRA on Unsloth, specifically trained for coding, agentic tool use, and reasoning, incorporating an anti-hallucination instruction set. Designed for local deployment, Rudra runs efficiently on CPUs or small GPUs via Ollama/llama.cpp, making it ideal for local coding assistants, function-calling experiments, and agent prototypes.

Key Capabilities

• Function Calling: Supports JSON-formatted tool calls for interacting with external functions like list_dir, read_file, write_file, edit_file, run_command, and search_code.
• Agentic Loop: Capable of planning, inspecting, editing, running, verifying, and summarizing tasks.
• Reasoning: Incorporates [VERIFY] and CONFIDENCE: markers to enhance reasoning and reduce hallucination.
• Local Inference: Provided in GGUF Q4_K_M format (~2.6 GB) for CPU-friendly operation, and compatible with Transformers.
• Improved Code Generation: In small-sample tests, rudra.Q4_K_M.gguf achieved 100% pass@1 on HumanEval+ (n=5), surpassing the base model's 60%.

Good for

• Local Coding Assistants: Its small size and optimization for code make it suitable for on-device programming support.
• Function-Calling Experiments: Ideal for developing and testing agentic workflows that require tool interaction.
• Agent Prototypes: Provides a capable base for building and iterating on AI agents with reasoning and anti-hallucination features.
• Resource-Constrained Environments: Runs effectively on hardware with limited GPU memory or even CPU-only setups.