GPT-OSS Turns One: Developer Shares Optimized Jinja Template

arbv · reddit · 2026-08-05

Celebrating the one-year anniversary of OpenAI's GPT-OSS models, a developer shared an improved Jinja prompt template. The author highly praised the GPT-OSS 120B as one of the best local models ever released, arguing it only faces real competition from Qwen 3.5 122B, while calling NVIDIA's Nemotron 3 Super a slower clone.

The shared template is based on the Unsloth version to ensure tool calls work correctly, incorporating TypeError fixes, additional sanity checks, and configurable token smuggling protection.

Original post →

More from Models

Models channel →