How I Built a 13B Parameter Model for Under $4K - Riemann Computing's "My Name is Riemann" Prerelease

The AI industry has become remarkably good at scaling infrastructure. My focus is different: scaling the learning process itself. More hardware can support larger training runs, but my ambition is to build a training approach that makes better use of the learning already accomplished.

That ambition led me to progressive model growth, NTK-inspired methods, and gating. Instead of treating every increase in model size as a reason to rebuild the training process, I wanted to explore a more deliberate path: preserve progress, introduce capacity in stages, and strengthen the model as it grows.

A Different Starting Point

My frustration was not simply with the cost of training. It was with the overhead surrounding it: repeated adjustments, memory-management complexity, and workflows that make experimentation harder than it needs to be.

In my own work, that meant questioning the role of LoRA-based workflows rather than assuming they were always the right fit. My goal was not to dismiss weight optimization. It was to investigate whether a different growth strategy could make the overall process more coherent.

The question became straightforward: what if learning could accumulate across planned growth stages, rather than each expansion becoming another disruptive transition?

Preserve Progress, Then Expand

The central idea is to save and stage learning so that model development becomes a sequence of connected phases.

My intended workflow is to:

• Preserve useful progress from earlier training stages.
• Introduce additional model capacity through controlled transitions.
• Use gating to manage how newly introduced capacity participates.
• Evaluate and stabilize the model before advancing further.

I think of this as hardening a model through growth. The objective is not merely to increase parameter count, but to make each expansion build meaningfully on what came before.

That distinction matters to me. A larger model is not the entire goal. A more intentional development process is.

Introducing My Name Is Riemann

That work has now reached a concrete milestone: the My Name Is Riemann Prerelease. The repository includes model weights, configuration files, and staged NTK controller artifacts. It turns this direction into something people can inspect and explore. You can explore the accompanying NTK Trainer and follow OpenPeer AI on Hugging Face as the work develops.

The model takes its name from my startup, Riemann Computing, which is behind the OpenPeer AI family of models.

My larger ambition is to help democratize the creator economy. I want independent developers, researchers, and creators to have more room to experiment, adapt, and build.

For me, that starts with challenging how models grow, not just celebrating how large they become. This release is my first public step toward making that ambition tangible.

Here are the results so far: