Non finito treats model quality as something you measure rather than assume. The core loop is simple: write a prompt, select two or more models, and read their answers next to each other. Public evaluations on the site span reasoning puzzles such as finding the least odd prime factor of a large number, visual deductive reasoning and real-world QA, RPM-style visual analogies, count-the-cards tasks, graph comprehension and policy recommendation, entity tracking, and base64 decoding — with models including OpenAI o1, DeepSeek R1, Qwen QwQ 32B Preview and xAI Grok 2 compared on identical inputs.
Beyond raw comparison, the platform records a verdict for each answer. Evaluations can be marked as pass or fail, which turns a pile of transcripts into something closer to a scoreboard you can revisit. That makes it useful for tracking whether a new model release actually improves on the previous one for your specific prompts, rather than trusting a leaderboard built from someone else's tasks.
Evaluations can be kept private in your workspace or published publicly for others to view and compare, and the site invites visitors to sign up and create their own. It is a lightweight research tool, so expect a focused interface rather than a full enterprise evaluation suite with automated scoring pipelines.
