Modelos pequeños en local que superan a Opus 🚀 Decision 3 - con David - @welcometolasecta 02x05 #041

Summary

The discussion focuses on the growing usefulness of small, locally run AI models, especially when they are specialized for particular tasks.

Model releases and development

The speakers describe Decision 3, a new open-weight model family associated with the vLLM team. It offers models ranging from roughly 600 million to 27 billion parameters and supports both text and vision. The range provides options for different hardware budgets. References in the subtitles to “trillions” appear to be transcription errors.

Decision 1.0, 2.0, and 3.0 reportedly appeared in quick succession, illustrating the fast pace of model development. The speakers also note that companies’ technical advantages can be copied or surpassed quickly.

Specialized models for fraud detection

The speakers discuss a comparison between a fine-tuned Qwen 3.5 9B model and Claude Opus 4.5 on fraud detection. The example illustrates how a smaller model trained for a narrow task can compete with—or outperform—a much larger, general-purpose model.

Why local models matter

Fraud detection and transaction analysis benefit from low latency. Keeping sensitive financial data local can also avoid sending it to an external service. The speakers expect capable hardware and models to make local AI practical for more workloads, while frontier models remain useful for tasks where they offer particular advantages.

Use at Maldita.es

David says Maldita.es uses frontier models for development and local models for information processing. He mentions Gemma and Qwen Next Flash, along with fine-tuning models for specific jobs. Some models are uncensored or have fewer guardrails, which can help with analyzing sensitive topics.

Latency and deployment

The organization aims to use fast, task-specific models for tasks such as transaction analysis and providing timely responses to users. Fine-tuning, combining models, and using an orchestrator are presented as ways to improve the overall system. Smaller Gemma models may also be suitable for mobile devices.

Reviews, guides, or tutorials

None are presented; this is a discussion of model releases and practical use cases.

Main speakers or sources

The podcast hosts, including Maxi, and guest David, who describes AI use at Maldita.es.