Company
Capability is not the finish line.
Geocentric is an open-weight intelligence and model engineering lab. We train models, build the evaluations that characterise them, and release them as open weights to run locally.
Evaluation is engineering work.
Geocentric works across the full model lifecycle: architecture, pretraining, post-training, evaluation, inference, vision, and release engineering.
A model earns trust to the extent it is understood. That means measurement before claims, published failures alongside successes, and evaluation treated as a first-class part of the training loop.
Engineering discipline is part of the product.
Measure before scaling
Profile speed, memory, context use, and model quality before turning an experiment into a claim.
Keep the architecture practical
Training tricks should not force exotic inference formats unless the gain justifies the complexity.
Publish the failures
An evaluation that only reports wins is marketing. Ours report both.
Own the full loop
Model quality depends on data, architecture, kernels, evaluation, post-training, and inference working together.