AIIA Qwen2.5-Coder Agent 7B

Qwen2.5-Coder-7B-Instruct fine-tuned for the AIIA agent framework — an
autonomous PLAN/BUILD-mode code agent with XML-style tool calling.

What it does

Executes coding and file-system tasks step by step using AIIA’s tool
protocol:

<WriteFile>                                                                                                                                                       
  <fileName>fib.py</fileName>                                                                                                                                     
  <contentOfFile>def fibonacci(n): ...</contentOfFile>                                                                                                            
</WriteFile>                                                                                                                                                      
                                                                                                                                                                  
<nextTask>completed</nextTask>                                                                                                                                    

Supported tools include: ReadFile, WriteFile, CreateFile, AppendFile,
TreeView, List, Find, Grep, ReplaceLine, Sed, Diff, ExecuteScript, Terminal,
createPlan, nextTask, and more. Task lifecycle is signalled with
<nextTask>completed</nextTask> / <nextTask>blocked</nextTask>.

Training

• Base: Qwen/Qwen2.5-Coder-7B-Instruct (Apache 2.0)
  
• Data: 2,296 real AIIA agent trajectories (file I/O, shell execution, code editing, planning, multi-step task completion) in ChatML format
• Method: QLoRA (4-bit NF4), r=32/alpha=64, loss on assistant content only, 3 epochs, final loss ~0.50
• Quantization: Q4_K_M GGUF (4.4 GB)

Usage

Designed to run inside the AIIA framework, which injects real tool results and drives the task loop.

ollama run w4d4f4k/qwen25-coder-aiia

Recommended options: temperature 0.3-0.7 for BUILD mode.

Compared to the other AIIA models

Use this one when coding quality matters most; use the smaller ones for lighter tasks and lower power draw.

Limitations

• Specialized for AIIA’s XML tool format; not a general chat model.
• Tool results must be provided by the calling framework; the model does not execute anything itself.
• English-focused, matching the training trajectories.

Requirements

• ~5 GB VRAM/RAM for Q4_K_M + context
• Runs on stock Ollama, llama.cpp, LM Studio — no custom runtime needed

Source

Merged bf16 weights: m5it/qwen25-coder-aiia-v1 (Hugging Face, private).