Open Source Hub

Lightweight local JavaScript model for efficient AI computations directly in your browser. Small, fast, and easy to integrate.

Enhanced version with improved performance and accuracy. More reliable outputs while maintaining efficiency for quick AI responses.

Universal model format for extreme portability. Run AI models on any device with zero dependencies—from servers to browsers to embedded systems.

Pre-computed template-based model for generating code, essays, and research papers. Fast, efficient, and pattern-driven responses without requiring heavy computation.

Open source conversational AI model trained on philosophical and deep knowledge corpus. Provides thoughtful, nuanced responses to complex questions.

Safety-first AI model built on pure training data. A recurrent neural network designed to be fundamentally safe and aligned. Runs entirely in your browser with no external API calls.

Enhanced safety-first AI model with improved training using Adam optimizer. Better accuracy and coherence while maintaining fundamental safety principles. 179-word vocabulary, 62.5K parameters. Trained with aligned data for safe responses.

Most advanced Lume version with improved tokenization (punctuation handling), larger 256-hidden-unit architecture, expanded training corpus with general knowledge, and robust safety guardrails. Better understanding of context and more coherent responses.

Production-ready model with batched training, 512 hidden units, mathematical reasoning capabilities with chain-of-thought, and critical bug fix for logit calculation. Trained on procedurally generated math problems. Handles complex BEDMAS expressions with step-by-step reasoning.

Enhanced version with temperature-based sampling for controlled creativity. Features 256 hidden units (optimized from 512), probability distribution sampling instead of greedy decoding, and configurable temperature (0.1-2.0) for balancing determinism and creativity. Includes batched training, mathematical reasoning with BEDMAS, and all critical bug fixes.

Replacing attention with frequency-domain global convolution for O(n log n) long-context modeling.

Revolutionary file format designed to democratize AI through extreme efficiency and universal portability. Run powerful models on any device with zero dependencies.

How AI can think in vectors, numbers, and invented languages more efficient than English—exploring higher-bandwidth machine cognition.

Comprehensive documentation of Agent-7 development: architecture, LLM challenges (instruction-following, format violations), and practical solutions.

AgAgent: An Autonomous AI Agent for Desktop Environment Interaction

Revolutionary approach to generating infinite high-quality training data through procedural mathematical expressions to improve AI reasoning.

Exploring the challenges of generalization, computational costs, and the need for more efficient algorithms in AGI development.

A revolutionary AI model designed to be fundamentally safe through pure training data. Lume is incapable of producing harmful content while maintaining strong reasoning capabilities.

AI's ability to perform research and development tasks, including model training, data analysis, and algorithm optimization for AGI.