yadavkapil7155/apex-qwen2.5

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:1.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 29, 2026License:mitArchitecture:Transformer Open Weights Featherless Exclusive Cold

Apex-Coder is a 1.5 billion parameter causal language model developed by Kapil Yadav, fine-tuned from Qwen2.5-Coder-1.5B-Instruct. Specialized for competitive programming, it excels at generating C++17 solutions for Codeforces-style problems. This model focuses on understanding problem statements, developing algorithms, and providing efficient C++ implementations with complexity analysis.

Loading preview...

Apex-Coder: Specialized for Competitive Programming

Apex-Coder is a 1.5 billion parameter model, fine-tuned by Kapil Yadav from the Qwen2.5-Coder-1.5B-Instruct base model. Its primary focus is on competitive programming, particularly for solving Codeforces-style problems and generating C++17 code. The model was fine-tuned using Supervised Fine-Tuning (SFT) with LoRA/PEFT and Unsloth, optimizing its ability to handle complex algorithmic challenges.

Key Capabilities

  • Competitive Programming: Designed to understand and solve problems typical of platforms like Codeforces.
  • C++17 Code Generation: Generates complete and efficient C++17 implementations.
  • Algorithmic Reasoning: Capable of identifying suitable algorithms, data structures, and solution approaches.
  • Complexity Analysis: Provides explanations for time and space complexity.
  • Debugging & Optimization: Can assist in debugging and optimizing C++ solutions.

Good For

  • Competitive Programmers: As an AI assistant for problem-solving and learning.
  • Educational Platforms: For teaching algorithms, data structures, and C++ coding.
  • Personal AI Coding Tools: Building lightweight assistants for C++ development.

Limitations

As a 1.5B parameter model, Apex-Coder may struggle with extremely difficult algorithms, complex dynamic programming, or problems requiring extensive multi-step reasoning. Users should always compile, test, and validate generated solutions, as the model is an assistant, not a guaranteed solver.