Ollama Ignored My Modelfile TEMPLATE. A Pre-Registered Token Count Caught It.
AI-generated summary A local benchmark run using Ollama and Qwen3.8 27B GGUF revealed that server prompt tokens remained constant despite changes to the Modelfile TEMPLATE. The author notes that Ollama 0.32.14 appears to render requests for models with the qwen35 family through a path other than the Modelfile TEMPLATE.