theoracle/gameagent-qwen2.5-coder-14b-merged

TEXT GENERATIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:14.8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 28, 2026License:otherArchitecture:Transformer Featherless Exclusive Cold

Theoracle/gameagent-qwen2.5-coder-14b-merged is a 14.8 billion parameter language model fine-tuned from Qwen2.5-Coder-14B-Instruct. It is specifically optimized to function as a smolagents.CodeAgent-compatible assistant for generating Pygame game code through an iterative Thought/Code/Observation loop. This model excels at writing simple 2D games like Snake, Catcher, and Pong, leveraging its specialized training on agentic trajectories.

Loading preview...

Model Overview

theoracle/gameagent-qwen2.5-coder-14b-merged is a 14.8 billion parameter model derived from Qwen2.5-Coder-14B-Instruct. It has been fine-tuned using LoRA (rank 16) to specialize in generating Python code for Pygame, specifically within the smolagents.CodeAgent framework. The model's training utilized real agentic trajectories from the theoracle/gameagent-pygame-trajectories dataset, focusing on simple 2D games such as Snake, Catcher, and Pong.

Key Capabilities

  • Pygame Code Generation: Designed to produce functional Pygame code for basic 2D games.
  • Agentic Workflow Integration: Optimized for an iterative Thought/Code/Observation loop, making it suitable for automated code generation and refinement.
  • Merged Weights: This repository contains the full-precision (bf16) merged weights, allowing direct loading with transformers without requiring peft.

Usage and Limitations

This model is best utilized within the smolagents.CodeAgent framework, where it can leverage its specialized training for agentic formats. It was trained with an 8192-token context cap, and quality may degrade in longer multi-step conversations. Known limitations include an occasional bug where it escapes the final-answer join as '\n'.join(lines) instead of ' '.join(lines). While effective for simple games, more complex tasks can exceed its optimal context length. The full-precision bf16 version is recommended over quantized versions due to observed quality loss in the latter.