Codaipro Files

CodaiPro v2.1.2

New recommended models for weak lab PCs, and cleaner answers from thinking models.

New

• Four sub-1 GB models join the recommended list and run on almost any lab PC: Qwen3.5-0.8B (newest), Qwen3-0.6B, Gemma 3 1B Instruct, Llama 3.2 1B Instruct — with direct download links on the website and in DOWNLOAD_MODEL.bat
• Thinking models (Qwen3 / Qwen3.5) no longer show their scratchpad: the <think> block is stripped and you get the answer. If the model runs out of tokens mid-thought, you get a clear hint instead of an empty reply

Fixed in v2.1.1 (recap)

• Fresh installs work again: all backend dependencies are declared
• AI engine upgraded to llama-cpp-python 0.3.x; every model gets its own chat template
• Single-instance guard watches the real backend port (8765); lockfiles are cleaned on exit; no duplicate backend spawn
• Generation timeout raised to 300 s; optional GPU offload via CODAIPRO_GPU_LAYERS

First run

1. Download a model into the models folder — sub-1 GB picks and direct links: https://luckyyaduvanshiofficial.github.io/Codai/#download
2. Run CodaiPro_v21.exe

System requirements

• Windows 10/11 (64-bit)
• 4 GB RAM minimum
• 2 GB free storage + 0.5–5 GB for one model

Full changelog: https://github.com/Luckyyaduvanshiofficial/Codai/compare/git-tag-v2.1.0...v2.1.2