kairos-sophron-1.0-14B

<p align="center"> <img src="kairos-logo.png" alt="Kairos" width="720"> </p>

Kairos Sophron 1.0 14B

Specialist local model for software engineering: smallest change that meets the requirement, clean-code boundaries, and no mix of refactor and feature in one change. Fine-tune of Qwen/Qwen2.5-Coder-14B-Instruct. This repository is the shipped Q4_K_M GGUF, the quant that was judged and deploy-tested.

The weights are a LoRA merged into bf16, then quantized to Q4_K_M (4.87 BPW, 8.4 GB). On a 12 GB GPU it stays on the GPU at num_ctx 8192, about 10 GB VRAM and ~34 tok/s. The file is already 8.4 GB, so an 8 GB GPU cannot hold the weights and the context together.

Gate

Both judges scored the adapter, and the deploy test scored this GGUF.

Zero case-level disagreements between the judges. Median token Jaccard of this GGUF against the judged adapter: 0.947. Replies in the user's language (EN or PT). Code identifiers stay in English.

Use with Ollama

Download kairos-sophron-1.0-14B-Q4_K_M.gguf and Modelfile into the same folder, then:

ollama create kairos-sophron:1.0 -f Modelfile
ollama run kairos-sophron:1.0

The Modelfile sets ChatML stops, num_ctx 8192, and temperature 0.

License

Apache 2.0, same as the Qwen2.5-Coder base. This repository does not include the training set.