Riemann Computing Details How It Built a 13.1 Billion-Parameter AI Model

Riemann Computing Inc. detailed how it expanded a 163 million-parameter language model into a 13.1 billion-parameter prerelease, outlining the hardware and evaluation results behind its staged development.

The My Name is Riemann model began with the published weights of OpenPeerLLM. Its development proceeded through successive stages rather than starting with a new model, reaching about 13.1 billion parameters in the current prerelease. A technical account of the process was published Oct. 28.

Initial development took place on a laptop with an external GPU. Later work moved to a cloud computing node equipped with eight AMD Instinct MI325X GPUs. The final development stage used about 2.16 billion training tokens.

OpenPeer-AI-NTK-Trainer used a neural tangent kernel-based controller to alter the model’s outputs while leaving its base weights intact. The approach supported experimentation across the successive development stages.

A held-out set of 1,000 prompt-and-completion pairs was used for evaluation. Applying the stage-six controller reduced negative log-likelihood from 1.323 to 1.310, while token accuracy increased from 67.73% to 67.84%.

Those results apply to the evaluation data and do not establish general-purpose or frontier-level AI capabilities. The prerelease needs broader evaluation, additional training data and improvements to the training process. It is an early development release for experimentation, not a finished consumer chatbot.

“Our vision is to create an ecosystem rather than being the sole authority in the AI revolution,” said Andrew Magdy Kamal, Riemann Computing’s founder.