Urgent.News

What's breaking now, across thousands of outlets.

AI

I ran a local vision model on my 8GB GPU against Claude and Gemini, and it made me rethink paying for either

Free vision is better than its reputation

I ran a local vision model on my 8GB GPU against Claude and Gemini, and it made me rethink paying for either

Running a local vision model on an 8GB GPU against large-scale options like Claude and Gemini led to a surprising realization – that smaller models have caught up, at least in theory. Empero AI's Qwythos 9B, fine-tuned from Alibaba's Qwen 3.5 9B, demonstrates this capability. Despite being trained primarily on text, it inherits Qwen's vision capabilities, allowing it to handle tasks like OCR and locating specific elements within images.

However, the vision aspect is not its strongest suit. Qwythos successfully extracted text from screenshots, matching Claude's performance, and even outperformed Gemini in some cases. For tasks primarily involving text-based screenshots, such as extracting a chatbot reply or file names, Qwythos proved to be a reliable alternative to larger, more expensive models.

However, when faced with more complex visuals, especially in decision-making scenarios, Qwythos fell short compared to models like Claude's Opus. While Qwythos offers a cost-effective solution for private, text-heavy tasks, it may not be the best choice for complex visual interpretation or when a more thorough analysis is required.

The author has decided to continue using Qwythos for private work, while reserving Claude and Gemini for more demanding tasks, questioning the value of his Gemini subscription in light of Qwythos's capabilities.

Written by urgent.news from XDA Developers's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.

Read the original at xda-developers.com →

More in AI

More from Thursday 19 November →