Miladasghari/light-coder

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 17, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Miladasghari/light-coder is an ultra-lightweight, instruction-tuned causal language model with 0.5 billion parameters, developed by Milad Asghari. Fine-tuned from Qwen/Qwen2.5-0.5B-Instruct, it specializes in code generation, refactoring, and following programming instructions. This model is designed for efficient execution on consumer hardware, requiring minimal VRAM and integrating seamlessly with standard inference tools.

Loading preview...

Overview

Miladasghari/light-coder is a 0.5 billion parameter instruction-tuned coding model, developed by Milad Asghari. It is built upon the Qwen/Qwen2.5-0.5B-Instruct base model, fine-tuned using approximately 122,000 programming instruction-response pairs. The LoRA weights were merged directly into the base checkpoint, resulting in a standalone model.

Key Capabilities

  • Ultra-lightweight: At under 1 GB in size (988 MB safetensors), it requires minimal VRAM.
  • Efficient Execution: Designed for quick execution on consumer GPUs and CPUs.
  • Code Generation: Primary domain includes generating code snippets.
  • Code Refactoring: Capable of assisting with code improvements and restructuring.
  • Instruction Following: Excels at understanding and executing programming-related instructions.
  • Seamless Integration: Works out-of-the-box with popular tools like vLLM, Ollama, and standard Hugging Face pipelines.

Good For

  • Developers needing a highly efficient, small-footprint model for local code assistance.
  • Applications requiring on-device code generation or refactoring where resources are limited.
  • Experimentation with instruction-tuned coding models without significant hardware investment.

Finetunes
1 models