Open-weight intelligence & model engineering lab
Models we can measure (before we trust them)
Geocentric is an open-weight intelligence and model engineering lab. We train language models, publish what our evaluations find, failures included, and release our models as open weights you download and run yourself.
Our flagship, announced and not publicly available.
ARCAvailable on Hugging Face and Ollama.
PARALLAXEvaluation across failure modes instead of one collapsed score.
EPICYCLETraining-system work across speed, context, token selection, and optimizer memory.
UNRELEASEDA further model in training.
Kestrel · Our flagship · Not publicly available
When the stars align.
Kestrel is built for Linux, C and networking, and it can spot flaws in code and systems. It learned from 57 programming languages, the entire Linux kernel and more than 80,000 man pages, on a single RTX 2060. It is announced, not publicly available. Because of security and safety concerns, we are not releasing it.
5,720,702,976 tokens. 249,396,480 parameters. Trained with our EPICYCLE system, which cuts the memory needed to train it by 44%.
How it was trainedArc · Now online
One model published. (Nothing hidden.)
Arc is a small English co-writing model. Download it and run it locally. Its model card publishes 24 unedited responses, including the ones it gets wrong.
Experimental. Fast local generation, with substantial accuracy and instruction-following limitations. The evaluation includes both successes and failures.
View the complete resultsMeasure the systems (not the story)
Geocentric benchmarks the pieces that decide whether a claim survives contact with hardware: generation speed, training throughput, and memory use. Small controlled tests give you numbers you can compare, repeat, and check.
higher generation throughput
74.26 → 88.27 tokens/sec in a controlled Apple M4 inference comparison after bounded KV storage and sampling-path changes.
M4 generation testless optimizer-state memory
Measured for experimental balanced momentum in an embedding-heavy 11.2M-parameter configuration.
Optimizer state / controlled testfaster context-folding pass
Context folding reached about 1.10× throughput in the M4 systems ablation at the same measured token count.
MPS training ablationSpend 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.
Grow 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.
Geocentric Code
Run it yourself (on your own machine)
Geocentric Code is a coding agent for your terminal. The model runs locally through Ollama, so your code and your prompts stay on your computer. One command installs everything.
Model identity
The model has a face
Geocentric's earth-themed character belongs to the model identity. The same visual language carries through the model family: grounded, technical, and a little unconventional.
