Ollama: set up Nimble and compare it with Jev

The tutorial covers decision models in Ollama for email routing, refund detection and urgency scoring. These calls return probabilities over typed answers rather than generated prose. The speaker explains binary probabilities, choices and scores, including why a 0.9 probability of yes does not measure intensity.

Setup uses Nimble and the /v1/system1 endpoint, followed by instructions for pointing the Typesafe SDK at a local server with a placeholder API key. Nimble has 9 billion parameters; the tutorial also discusses TEV 1 variants at 4B and 0.8B. This is a way to run models locally for classification workflows.

On an M5 Max with 128 GB RAM, the speaker reports 70 ms for Nimble and 50 ms for TEV 1. The hosted comparison includes a round trip from Korea, so the figures do not establish a universal speed advantage. Cited accuracy tests favor Nimble on some classification tasks, while hosted Jev leads on a reasoning panel.

Limits matter: choice and score questions accept up to 26 options, with 2 to 24 recommended for TEV 1. Its roughly 2,000-token budget includes the rendered prompt; Nimble has an 8,192-token context. Oversized prompts fail rather than truncate. The speaker also criticizes hosted retention terms and argues that fixed outputs reduce prompt injection risk. Those are the speaker's assessments, not an independent privacy or security audit.