Nous Research Launches Forge API Beta to Boost AI Reasoning Power

Nous Research has announced the beta launch of its Forge Reasoning API, designed to elevate the reasoning capabilities of large language models (LLMs) like Hermes 3 70B. Forge combines advanced reasoning architectures—such as Monte Carlo Tree Search (MCTS), Chain of Code (CoC), and Mixture of Agents (MoA)—to enhance LLMs, enabling them to handle complex real-world scenarios with precision.

The company has also introduced Nous Chat, a streamlined interface for using the Hermes 3, an open-source model. The threaded conversation system helps organize your thoughts and projects, while customizable prompts and settings give you full control over AI interactions. Available at hermes.nousresearch.com, the service is currently free.

Hermes 3 70B with Forge: Competitive with Top AI Models

The Forge Reasoning API lets you enhance any popular model with a code interpreter and advanced reasoning abilities.

Forge’s architecture has demonstrated impressive performance enhancements for Hermes 3 70B, a 70-billion-parameter open-source LLM. With Forge’s enhancements, Hermes 3 70B achieves competitive scores against industry-leading models like GPT-4, Gemini, and o1 in key benchmarks:

• AIME (competition-grade math): Hermes 3 70B x Forge achieves an 80.00% score, surpassing many larger models.

• MMLU-Pro (broad knowledge test): The model scores 67.50%, demonstrating its versatility.

• MATH: Hermes 3 70B x Forge performs well at 81.30%, suitable for complex math questions.

This success, particularly in the AIME test used for math competitions, shows Forge’s potential to enhance model performance in high-stakes, real-world applications.

Introducing the Forge Reasoning System

Forge integrates three powerful reasoning systems to push the boundaries of LLM inference:

• Monte Carlo Tree Search (MCTS): Useful in planning problems, MCTS builds decision trees to simulate and evaluate various actions, optimizing the LLM's decision-making process.

• Chain of Code (CoC): CoC allows the model to connect reasoning steps with a code interpreter, enhancing its ability to solve math and code-based questions that require sequential problem-solving.

• Mixture of Agents (MoA): MoA enables multiple models to work together, combining their outputs for a more nuanced response than any single model can provide. It fosters collaboration between models, improving response quality.

Enhanced Flexibility and Real-World Applications

The Forge Reasoning API is designed with versatility, supporting multiple models (Hermes 3, Claude Sonnet 3.5, Gemini, GPT-4), and allowing users to choose a single model or combine models for diverse outputs. This flexibility ensures that Forge can adapt to various use cases, from research applications to field-specific problem-solving.

Nous Research is rolling out the API to a select beta group, focusing on gathering feedback to refine Forge’s capabilities. Lambda, Nous Research’s compute partner, is providing resources to support this beta testing phase.

Demonstrating Nuanced Output: Forge vs. o1 Preview

A comparison of Forge and o1 model outputs illustrates Forge's capacity for nuanced, creative responses. In scenarios requiring imaginative language, such as role-play, Forge’s multiple-agent architecture enables varied and flexible responses, giving users more choice and control over the output.

• Forge’s Strengths: It offers nuanced, flexible, and imaginative outputs, which are advantageous for complex and creative tasks.

• Comparison to o1: Forge appears more adaptable in generating varied responses that users can choose from, which could be seen as a "better" feature in contexts where response diversity and user preference matter.

In real-world applications, Forge could therefore be preferred in tasks where customization, multiple perspectives, or creative reasoning are key.

Looking Forward: Elevating LLM Technology with Forge

The Forge Reasoning API represents a leap forward in LLM reasoning and inference, bringing closer the vision of more autonomous, intelligent AI systems. As Nous Research gathers user feedback from the beta program, they aim to expand Forge’s capabilities, empowering developers and researchers with a tool for advanced reasoning tasks that support real-world complexity.

Editor’s Note: This article was created by Alicia Shapiro, CMO of AiNews.com, with writing, image, and idea-generation support from ChatGPT, an AI assistant. However, the final perspective and editorial choices are solely Alicia Shapiro’s. Special thanks to ChatGPT for assistance with research and editorial support in crafting this article.