sandeeprdy1729/TIMPS-Coder-0.5B

TEXT GENERATIONConcurrent Unit Cost:1Model Size:0.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Apr 12, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

TIMPS-Coder-0.5B by Sandeep Reddy is a 0.5 billion parameter Qwen2 Transformer model, fine-tuned for specialized code-related tasks. With a 4096-token context window, it excels in bug analysis, code review, algorithm problem-solving, and agentic planning. This model is optimized to "think before it codes," achieving an 88% internal benchmark score across various software engineering dimensions.

Loading preview...

TIMPS-Coder-0.5B: Elite Bug-Fixing Assistant

TIMPS-Coder-0.5B, developed by Sandeep Reddy, is a 0.5 billion parameter Qwen2 Transformer model built on Qwen/Qwen2.5-Coder-0.5B-Instruct. It is specifically fine-tuned using LoRA to enhance its capabilities in code-related tasks, emphasizing a "think before it codes" approach. The model features a 4096-token context window and is available in Q4_K_M GGUF (Ollama) and BF16 safetensors (HuggingFace) formats.

Key Capabilities

  • Bug Root-Cause Analysis: Proficient in identifying and explaining the root causes of bugs, including SQL injection, race conditions, and memory leaks.
  • Code Review: Capable of reviewing code and detecting potential issues.
  • Algorithm Problem-Solving: Excels at LeetCode medium-level algorithm problems and O-notation analysis.
  • Agentic Planning: Supports agentic reasoning and planning, including tool use and bash planning.
  • Software Engineering Tasks: Achieved an 88% total score in internal benchmarks, with 90% in Bug Fix, SWE/Repo-level, and Algorithms, and 80% in Code Review.

Training Details

The model was fine-tuned using LoRA (rank=16) over 3,000 iterations on Apple M-series hardware. The training data comprised approximately 8,400 samples from diverse datasets, including LeetCode problems, real GitHub issues (SWE-bench), agentic edit traces, and a custom bug-fix corpus. All training samples were formatted with a THINK → FIX → VERIFY answer structure.

Usage Tips

For optimal performance, it is recommended to use a temperature of 0.1 to minimize hallucinations. Always include the full function or class when requesting bug fixes and verify generated code, as even with high accuracy, edge cases can exist. A system prompt, as shown in the Quick Start examples, is crucial for best results.