I ran a 9B model trained on Claude's reasoning on my 8GB GPU, and it made me rethink paying for Claude

I recently downloaded a free reasoning model called Qwythos 9B, created by Empero AI, and experimented with running it on my 8GB GPU. This model claims to have learned from Anthropic's Claude models, specifically Claude Mythos and Claude Fable, having been trained on over 500 million tokens of their conversations. The model's step-by-step reasoning process is facilitated by a separate block that thinks through a problem before responding, which is a feature I find valuable for tasks like explaining concepts or breaking down complex problems.

When I tested Qwythos on various tasks, it performed well, providing accurate answers quickly, although it sometimes struggled with specifics like coding and design tasks. It was particularly impressive in handling logic puzzles and scientific explanations, often providing detailed and accurate responses. While Qwythos can't access real-time information due to its offline nature, it does well with tasks that don't require the latest data, making it a useful tool for explaining concepts or breaking down problems.

Its performance on vision tasks was not as strong, likely due to the model not being trained or tested for these kinds of tasks. Overall, running Qwythos on my GPU has allowed me to save on Claude's subscription costs for smaller tasks, but I still rely on Claude for more complex or real-time tasks.