Introducing ZYR3 v1 — An Agentic Coding AI

A developer released ZYR3 v1, an agentic coding model published on Hugging Face as a LoRA/PEFT adapter built on the Qwen2.5-Coder-7B-Instruct base model in Safetensors format. ZYR3 v1 targets code generation, debugging, editing, refactoring, code explanation and multi-step coding tasks, and can run locally on NVIDIA GPUs or through inference frameworks including Transformers, PEFT, vLLM, SGLang and llama.cpp. The developer said a free API is in development to let users try ZYR3 from a browser or their own applications without downloading weights, with the endpoint to be announced later.

Introducing ZYR3 v1 — An Agentic Coding AI I’m excited to share ZYR3 v1 , my latest agentic coding model available on Hugging Face: Model: zyr-AGENT/zyr3-v1 ZYR3 v1 is built as a LoRA adapter on Qwen2.5-Coder-7B-Instruct , designed to behave more like a coding agent than a simple code completion model. • Code generation • Programming reasoning • Debugging & bug finding • Code editing • Refactoring • Code explanation • Iterative problem solving • Multi-step coding tasks • Edge-case handling The goal is simple: Don’t just generate code — work through the problem. Architecture: LoRA / PEFT adapter Base model: Qwen2.5-Coder-7B-Instruct Task: Text Generation / Agentic Coding Format: Safetensors Framework: Transformers + PEFT ZYR3 v1 can be run locally using modern NVIDIA GPUs, with quantized versions making it possible to run with significantly lower VRAM. You can integrate it into your own coding workflow using tools such as Transformers, PEFT, vLLM, SGLang, llama.cpp, and other compatible inference frameworks. We are also working on making ZYR3 v1 available through a free API . The goal is to let developers try ZYR3 directly from the browser or their own applications without installing the model or downloading any weights . For an initial period, the API is planned to be free to use , so anyone can experiment with ZYR3 and see what it can do. More details and the API endpoint will be announced when it’s ready. Most coding models focus primarily on producing code. ZYR3 is designed around a different workflow: Understand → Plan → Code → Debug → Refine → Repeat The goal is to make ZYR3 useful for real development workflows, not just isolated code-generation prompts. Try the model: zyr-AGENT/zyr3-v1 More updates coming soon. AI LLM Coding CodingAgent AgenticAI HuggingFace Qwen OpenSource MachineLearning ArtificialIntelligence