Mostik Bridges 753 Billion- and 4 Billion-Parameter AI Models at 1/20 the Cost
Updated
Updated · WIRED · Sep 2
Mostik Bridges 753 Billion- and 4 Billion-Parameter AI Models at 1/20 the Cost
1 articles · Updated · WIRED · Sep 2
Summary
Mostik says it can make AI models communicate through weight values rather than text, letting a smaller model absorb part of a larger model’s capability more efficiently.
In a demo, the startup linked a 753 billion-parameter GLM-5.2 model with a 4 billion-parameter Qwen-3.5 model, producing a hybrid that costs one-twentieth as much as full GLM and performs midway between the two.
The company says the approach also powered a model now leading ARC-AGI 3, though it withheld technical details while trying to win the competition.
Researchers familiar with the work say the method could help open-weight and domain-specific models approach frontier-model quality without running a giant model through the full inference loop.
Mostik argues that if model-bridging scales, AI progress may come less from ever-larger monolithic systems and more from networks of specialized models working together.