Prime Intellect supports full-parameter training
Prime Intellect
Moving from LoRA to full parameter training shifts Prime Intellect from a lightweight customization tool into a true model building service. With LoRA, customers only trained a small add on file that sat on top of an unchanged base model, which kept jobs cheap, fast, and easy to share. Full training means the core model itself changes, which is what teams need when they want to bake in new behavior, domain knowledge, or architecture level improvements instead of just layering on a narrow task specific patch.1 2 3 4
- LoRA is designed to update a small set of adapter weights while freezing the original model. That keeps checkpoints small and memory use low, and it makes one base model easy to reuse across many task specific variants. It is a good fit for cheap fine tuning, not for rewriting the model itself.2 3
- Prime Intellect now supports hosted full parameter training on Prime managed GPUs, sold either on demand by GPU hour or through a dedicated cluster. That changes the product from shared adapter training into heavier infrastructure, where customers are effectively renting a training factory to produce their own model checkpoints.1 4
- This also puts Prime Intellect closer to the workflow of serious post training platforms. Thinking Machines Tinker still centers managed LoRA fine tuning with downloadable checkpoints, while Prime Intellect is moving up the stack into jobs that consume more compute, take longer, and are harder for customers to run on fragmented external GPU supply by themselves.1 4 5
The next step is a tighter loop between Lab and Prime Agent. If Prime Intellect can train the whole model, then immediately drop that updated model into an agent runtime for coding and research work, it can sell not just cheaper GPU access but a full system for improving and deploying agent behavior over time.1 4