27B model with 5 skills outperforms expectations on healthcare workflows

Ubunta · x · 2026-08-28

The author experimented with enhancing a small 27B model for healthcare workflows by adding 5 specific skills (cohort specs, value sets, etc.) instead of upgrading to a larger model. The setup runs locally on 2 consumer GPUs without fine-tuning. All numerical outputs are retrieved via tool calls, not generated by the model. Initial observations suggest performance gains come primarily from the skill and context layer, though loading the wrong skill remains a challenge.

Original post →

More from coding & agent

coding & agent channel →