Konic Fields
Small models for deep fields.
Konic Fields is our research program for specialized small models. We train them deep on one domain, then make them small enough to run on the device, at the edge or inside your lab.
- Status
- Research program
- Now
- Choosing first partners
- Runs on
- Device, edge, your own lab
Depth beats width.
The hard problems in a field are specific. A small model trained deep on that field can compete with general models many times its size on the tasks that matter there.
Small enough to go where the work is.
On a device, an instrument, a robot, a lab workstation or a server at the edge. No data center and no dependable connection required.
Private by construction.
Training runs inside your environment and the model runs on your hardware. Your data never leaves.
Where we want to work
Open to any field.
These are the fields we want to start with. If you work on a hard problem in another one, we want to hear about it.
Life sciences and drug discovery
Molecules, assays and biology.
Healthcare and clinical
Notes and imaging reports, inside the hospital.
Materials and chemistry
Formulation, catalysts and batteries.
Semiconductors
Chip design and verification.
Engineering
Simulation and embedded systems.
Energy
Grid and generation data on isolated networks.
Industry
Plant, machine and sensor data at the line.
Defense
Sealed networks where nothing may leave.
Cybersecurity
Malware analysis and log triage, air-gapped.
Aerospace and space
On-board compute with no reliable link.
Robotics and autonomy
Multimodal models on the machine itself.
Geoscience and climate
Subsurface, weather and earth observation.
Telecom and networks
Network data that stays with the operator.
Finance and quant research
Proprietary signals and filings.
Agriculture and food science
Crop, soil and formulation data.
Your field
Bring us a hard problem and its data.
How it works
From your data to a model on your hardware.
Together we fix the tasks, the data and a benchmark your domain experts trust.
Konic's post-training pipeline runs on your infrastructure. Your data never leaves it.
Pruning, distillation and low-bit quantization fit the model to the device or server it will run on.
We measure it against far larger models on your benchmark and decide on the numbers.
A general model knows a little about everything.
Training on your data concentrates it on your field.
Compression removes what your field does not need.
What remains is measured against far larger models.
The research behind it
We already make models small.
Konic Fields has no model to announce yet. This is the published compression work it builds on.
4.0×
smaller weights at 1.034× baseline perplexity
From-scratch AWQ INT4 quantization
96.8%
of code-generation performance kept at 25% compression
REAP expert pruning
2.79 GB
packed artifact from an 8B-parameter model
Two-stage pruning and INT4
59.9%
smaller multimodal encoder, retaining 0.973 AUROC
INT4 encoder compression
First partners
What a first partner gets.
Early access
First use of each domain model, before any wider release.
Field exclusivity
No competing partner in your exact domain for an agreed period.
Joint benchmark or paper
A shared evaluation suite and co-authored publication.
Preferred licence terms
Better commercial terms on the models that result.
Ways to work together
Co-development
You bring the domain data and benchmarks. We bring post-training and compression.
Annual licence
A domain model licensed to run on your own infrastructure, with versioned updates.
Custom build
A model trained for your organization, on your data, as a scoped project.
Open weights
Selected domain models released openly, with enterprise licences and support.
Konic Fields
Bring us a hard problem,
we'll bring the small model.
We are choosing our first partners now, in any field with hard problems and its own data.