NSED: Mixture-of-Models Achieves SOTA Reasoning with Self-Hosted AI

The article introduces NSED (N-Way Self-Evaluating Deliberation), an enterprise-ready agentic orchestration layer that implements a mixture-of-models architecture. NSED uses multiple AI agents working in parallel, with intermediate checkpoints and convergence on a final answer. The key benefit is frontier reasoning ability achieved through amplifying the strengths of individual models. The system shows less sycophancy, offers cost tracking and optimization, and provides a full compliance audit trail managed through NATS. Remarkably, three small open-weight models (20B, 8B, 12B) on consumer hardware score 84% on AIME 2025 through NSED deliberation, matching DeepSeek-R1 and coming within 1 point of GPT-5, while naive majority voting only achieves 54%. NSED addresses several key challenges, including single-model quality ceilings, black-box AI decisions, data privacy, lack of human oversight, vendor lock-in, and high hardware costs. It runs entirely on local infrastructure, ensuring data never leaves the network, and it is provider-agnostic, allowing users to mix OpenAI, Ollama, vLLM, or any OpenAI-compatible endpoint in the same session. NSED achieves high reasoning quality with a swarm of smaller models, reducing the hardware requirements compared to monolithic models.

Transparency in AI decision-making is crucial for building trust and ensuring accountability. NSED provides a full audit trail of every proposal, evaluation, score, and reasoning trace, which can be streamed via SSE. This level of transparency allows users to understand how the system arrives at its conclusions and to identify any potential biases or errors.

Ultimately, NSED represents a significant step towards democratizing access to advanced AI capabilities. By enabling high-level reasoning on consumer hardware and ensuring data privacy, NSED empowers users to leverage the power of AI in a responsible and cost-effective manner.