Project Overview | IN DEVELOPMENT
GENESIS-100M is not a fine-tune of an existing model. It is a from-scratch pretrain -- 100 million parameters, trained entirely on data that our own AI systems curate and generate. No dependency on GPT, Claude, or any external model API. The training data is home-grown; the weights are our own; the architecture is our design.
Why 100 Million Parameters?
Conventional wisdom says bigger is better. We reject that premise. A well-trained 100M parameter model with high-quality synthetic data can outperform a 7B model trained on noisy internet text. The difference is not parameter count -- it is data quality.
When AI curates its own training data, it can achieve far higher information density per token. Instead of learning from Reddit comments and SEO spam, GENESIS-100M learns from carefully selected, verified, and synthesized content. Every training example is scored for quality before inclusion.
| Approach | From-scratch pretrain. No foundation model dependency. No fine-tuning on existing weights. |
| Data Source | 100% AI-curated -- Prometheus pipeline selects, filters, and synthesizes all training data. |
| Architecture | Decoder-only transformer with modern efficiency optimizations. Exact architecture TBD pending design review. |
| Infrastructure | Self-hosted, GPU-accelerated. No cloud dependency, no vendor lock-in, no recurring API costs. |
The Case for Small, Efficient Models
Large language models have become a commodity -- accessible via API but impossible to own, audit, or truly control. They run on hardware you don't control, trained on data you can't inspect, serving through APIs that can change or disappear overnight.
GENESIS-100M represents a different path: a model small enough to run on modest hardware, efficient enough to be practical, and transparent enough to be verifiable. When you can host your own model, you don't need a terms-of-service agreement -- you have actual ownership.
The future of AI belongs to efficient, verifiable models -- not bloated black boxes. GENESIS-100M is our first step on that path.
Timeline
| Q2 2026 | Architecture design & data pipeline completion (Prometheus Phases 1-3) |
| Q3 2026 | Initial pretraining run. First checkpoint evaluation. |
| Q4 2026 | Model release. Public benchmarks. Open-weight availability. |