We compress frontier language models until they fit on the hardware you already own.
The wall
A frontier open-weight model in its original precision needs more memory than any single accelerator has. So teams rent a multi-GPU node they cannot saturate, or they give up and run something smaller and worse. Memory is the wall, not compute.
What we do to it
The same 122 billion parameters, still generating text. And because we ship the kernels that read the format, that figure is what it occupies in GPU memory — not a packed file that expands the moment you load it.
Deployment cost · 01
Four accelerators to serve one copy becomes one accelerator serving two. That is eight times the served capacity per card — the same model, on a quarter of the silicon.
Deployment cost · 02
Both resident in VRAM at once, on one card you can actually buy, in a chassis you already have — instead of a rented multi-GPU node with an interconnect you are paying for and not using.
Deployment cost · 03
The same artifact runs on commodity CPU, and Parakon 30B runs on a 16 GB M3 MacBook Air through kernels we wrote ourselves. Batch work stops needing a GPU queue.
Feel the speed
Replayed, character for character, at the measured generation speed of Parakon 30B on a 16 GB M3 MacBook Air. No cloud, no GPU, no queue — this is what a 30-billion-parameter model feels like on a laptop.
On premises
Small enough to sit inside the perimeter you already defend. No inference provider, no egress, no per-token meter, no third party holding your prompts. Air-gapped, regulated and on-prem estates get a frontier-class model on hardware they own outright.
The method
Uniform compression treats a routing gate like a feed-forward expert. That is why it falls apart at the extremes, and why ours does not.
Allocation
Nearly nine tenths of the network is stored at the extreme end of the scale, while the handful of tensors that route every token keep full precision. One bad gate routes every token wrong.
Kernels
Our models ship with the software that executes them, so the small footprint holds in GPU memory instead of only on disk — on datacenter GPUs, on commodity CPU, and on Apple silicon, where nothing else runs models this size at all.
Fit
Scales are fitted against measured activation statistics, expert by expert, so precision follows where the model actually sends its tokens rather than a uniform prior.
Who runs Parakon
Enterprise
A fine-tuned model and a GPU invoice. Four cards become one, and the weights never leave the building.
Clouds & GPU networks
Every gigabyte freed is another paying workload on the same silicon. Compression is margin.
On-device
Frontier-class capability on a consumer laptop is the existence proof. Devices ship intelligence, not API keys.
Silicon
Chips live or die by what fits in their memory. We make the models the market wants fit the hardware being sold.
Evidence
Every figure is taken from the exported artifact on the runtime it ships with, never from a training checkpoint. Nothing is adopted on a single run. Where we have not measured something, we say so — including here.
Parakon 30B · full suite, retained
Each figure is the compressed model's score as a share of the same checkpoint at full precision, run through one harness at 10× less memory. Retention holds for knowledge, tool use and grade-school reasoning; it thins on code and instruction format, and we publish those two at the same size as the rest.
IQ per gigabyte
Points of retained benchmark quality per gigabyte the weights occupy in memory. Full precision scores 100% of itself by definition — and spends every gigabyte to do it.
Parakon 122B · in progress
Parakon 122B's suite is live on the bench right now. These are the benchmarks that have completed, published as they land rather than held back until the average flatters us.
Retention against the published reference
Knowledge and instruction following come back at better than nine tenths of the uncompressed model. Long-chain scientific reasoning is the one that costs us, and it is on the page at full size because that is the honest shape of the result.
Parakon
Parakon
122 billion parameters. One accelerator. Inside your own walls.
Weights are available under an evaluation license — run our numbers yourself before a single conversation about deployment.