Forge Boosts Small AI Model Performance

A new open-source tool called Forge uses guardrails to dramatically improve the reliability of small, self-hosted AI models on complex agentic tasks.

A new open-source tool named Forge has been released to improve the reliability of self-hosted large language models (LLMs). Developed by an AI Director at Texas Instruments, Forge acts as a reliability layer for local models running on consumer hardware. It introduces a set of 'guardrails'—including automated retries, error recovery, and context management—that operate around the model. The key finding is a dramatic performance increase on multi-step agentic tasks, with an 8-billion-parameter model jumping from a 53% success rate to approximately 99%. This improvement is achieved without any changes to the underlying model itself, focusing instead on strengthening the system that directs it.

This development is significant for developers and businesses building applications with smaller, locally-run AI models. By making these models more dependable, Forge lowers the barrier for creating sophisticated, always-on AI agents that can perform complex workflows without relying on larger, more expensive cloud-based services. The ability to achieve near-perfect reliability on smaller models makes advanced AI more accessible and cost-effective. The project also includes an evaluation framework and an interactive dashboard, allowing users to reproduce the performance claims and test the system with their own setups.