GENOMA LABS / KALYPSO: Open-Weights Coding and Agent Models

KALYPSO: Open-Weights Coding and Agent Models

KALYPSO is GENOMA Labs' model line for coding and agent orchestration. The public release, KALYPSO v1.1L, is a 15-billion-parameter model with open weights under Apache-2.0, available now on Hugging Face.

KALYPSO writes code, and it also directs the agents that write code. GENOMA builds models for specific applications, from robotic world models to engineering and simulation systems, and KALYPSO is the line that plans, delegates, and verifies that work.

v1.1L: 15 Billion Parameters, Apache-2.0

v1.1L is trained for two roles: generating code and coordinating other agents. The Apache-2.0 license means you can download the weights, run them on your own hardware, fine-tune them, and ship the result in commercial products. The weights are public at huggingface.co/genomalabs/KALYPSO-v1.1L.

Agent orchestration is a first-class capability, not a side effect. v1.1L can take a task, split it into independent units, dispatch those units to worker agents, and pull the results back together.

Deterministic Gates: Verify by Running

Every KALYPSO generation is checked by a deterministic gate before it ships. A gate is a concrete test, a command that exits zero on success. Output is not judged by another model or by prose review. It is run, and the run is the verdict.

  • Write the gate first. A precise spec plus a failing check defines done before generation starts.
  • Fan out. Independent units go to parallel KALYPSO agents, each unit with its own gate.
  • Ship only green. Output that never passes its gate does not ship, and a gate is never weakened to let it through.

v1.2L in Production, Served on GENOMA Hardware

v1.2L is the production serving line. Serving is sovereign and local-first: GENOMA trains and serves on hardware it owns, housed in GENOMA Pods, L-series compute capsules that link from a single rack to a full data hall. An internal K-line for robotics exists alongside the public L-line and is not published.

GENOMA applies the same approach across its products, including COPMA AI, its AI platform for the heavy-equipment industry.

KALYPSO v1.2L is public as well: a 31-billion-parameter model released with open weights under Apache 2.0 on Hugging Face. It is a stronger orchestrator and a stronger model for complex coding than v1.1L, and it is the base of the production serving line.

Questions

Can I use KALYPSO v1.1L commercially?

Yes. The weights are released under Apache-2.0, which permits commercial use, modification, and redistribution. Download them from Hugging Face and run them on your own hardware.

What is the difference between v1.1L and v1.2L?

v1.1L is the public open-weights release, available on Hugging Face under Apache-2.0. v1.2L is the production serving line GENOMA operates on hardware it owns. A separate K-line exists for internal work and is not published.

What does KALYPSO do beyond writing code?

It orchestrates agents. KALYPSO splits a task into independent units, fans them out to parallel agents, and checks every result against a deterministic gate before anything ships. The gate, not the model, decides what passes.