AI models built close to the hardware.
Geocentric builds compact language models and the systems that train them, with a focus on capability per parameter, fast iteration, and efficient use of hardware you can actually own.
A small co-writer.
Ready to write.
Arc explores low-cost English co-writing on local hardware. Start a conversation or inspect its model card and evaluation.
Experimental. Fast local generation, with substantial accuracy and instruction-following limitations. The evaluation includes both successes and failures.
View the complete resultsAPPLE M4
Measure the systems, not the story.
Geocentric benchmarks the pieces that determine whether local model work is actually practical: generation speed, training throughput, and memory use.
higher generation throughput
74.26 → 88.27 tokens/sec in a controlled Apple M4 inference comparison after bounded KV storage and sampling-path changes.
less optimizer-state memory
Measured for experimental balanced momentum in an embedding-heavy 11.2M-parameter configuration.
faster context-folding pass
Context folding reached about 1.10× throughput in the M4 systems ablation at the same measured token count.
Spend compute where it matters.
EPICYCLE is Geocentric's training-efficiency system for getting more useful work from constrained hardware without changing the model's ordinary inference architecture.
Read the technical breakdownGrow depth during training
Start shallow and activate later layers as the run develops.
Fold context before expanding it
Train on shorter independent segments before moving to full context.
Focus token loss
Experiment with selective token-loss bands after warmup.
Cut persistent optimizer state
Trade exact dense optimizer state for more model capacity.
The model has a face.
Geocentric's earth-themed character is part of the model identity, not a generic mascot pasted onto the company after the fact.
The same visual language carries through the model family: grounded, technical, a little unconventional, and unmistakably Geocentric.
See Arc, Meridian, and Zenith
LATEST WORK