Hyperion-DeepSpace-218M

Emergence Labs • 2025

"Reasoning is not emergent.
It is engineered in 64 parallel paths."

• 218 million parameters
• Trained on TinyStories in under 1 hour on a single GPU
• Final PPL: ~7.3 (better than most 1B+ models)
• No RLHF • No synthetic data • No preference tuning
• Pure next-token prediction with explicit symbolic decomposition

This is the first public release from Emergence Labs.

We are just getting started.

WebEssentz / Onyerikam Godwin Gideon
December 2025