The intelligence layer for machines that move the world.
GENOMA Labs is an AI lab that builds models for specific applications: robotics VLA and world models, engineering and deep-tech models, simulation and training systems. We train them, own them, serve them ourselves, and ship them as products in market.
01 / Models
The right model for each application, not one model for everything.
General models stop at the answer. We build application-specific models that close the loop to action, each trained for its domain, gated by deterministic checks, and served on infrastructure we own.
Frontier model APIs are expensive and restricted. Open-weight alternatives often carry provenance questions. And proprietary data should never leave systems you control. That is why we exist: models made for your needs, deployed on cloud compute or fully local, free of the restrictions of general-purpose frontier models.
Robotics Models
Vision-language-action and world models for embodied intelligence, with reinforcement learning for heavy machinery. Models that perceive a physical scene, predict its dynamics, and act in it.
- VLA models
- Robotic world models
- RL for heavy machinery
- Real-to-sim & sim-to-real
Engineering & Deep-Tech Models
Models tuned for engineering work and research: the KALYPSO coding and orchestration line, research models for deep-tech domains, and proprietary models kept in-house.
- Engineering models
- Research & deep-tech models
- Agentic orchestration
- Sovereign, local-first serving
Simulation & Learning Models
Simulation models and the training-learning systems around them: synthetic worlds to learn in, data acquisition loops to learn from, and deterministic gates that decide what ships.
- Simulation models
- Training-learning systems
- Data acquisition loops
- Deterministic quality gates
02 / Research
Where we push the frontier.
Embodied intelligence is a systems problem. Our research spans the model, the simulator, and the machine, because the hard part is making all three agree in real time.
03 / Flagship
KALYPSO: our frontier model line, in the open.
KALYPSO is the model family we train and serve ourselves, with open weights on Hugging Face. v1.1L is a 15-billion-parameter coding and agent-orchestration model. v1.2L raises it: 31 billion parameters, a stronger orchestrator, and a stronger model for complex coding, both under Apache 2.0.
Behind it sits the same discipline as our robotics stack: every generation is checked against a deterministic gate before it ships, and the serving path stays on infrastructure we own.
04 / Products
Research, shipped.
Every model line ships as a product. Open weights on Hugging Face today, robotic world models coming to fal.ai, and native AI products live in real estate, gaming, and heavy industry.
Craneformerz
Native AI · craneformers.ioAn AI videogame and data-acquisition platform for robotic lifting operations. Drive the machines in play; every session generates the operational data that trains the models.
craneformers.io →
COPMA AI
Live · copma.aiAn AI platform for the heavy-equipment industry: grounded product knowledge, parts and service intelligence, and multilingual support for a global crane manufacturer.
copma.ai →Enzo’s
Native AI · enzos.aiAI for real-estate search. Find real homes by architectural style, feature, and mood, with design-literate descriptions and a demand index no listing portal can match.
enzos.ai →KALYPSO v1.1L
Open weights · 15BThe public entry to our coding and orchestration line: 15 billion parameters, Apache 2.0. KALYPSO writes, verifies, and repairs software behind a deterministic gate, and fans out as a fleet of agents.
huggingface.co/genomalabs →KALYPSO v1.2L
Open weights · 31BOur strongest open release: a 31-billion-parameter model under Apache 2.0, a stronger orchestrator and a stronger model for complex coding, and the base of the production serving line.
kalypso-v1.2l on Hugging Face →Robotic World Models
Soon · fal.aiGenerative world models for embodied AI: physical scenes and machine dynamics, predicted frame by frame, for planning, simulation, and training. Launching on fal.ai.
Coming to fal.aiRL for Heavy Machinery
Applied researchReinforcement learning stacks for industrial autonomy: policies trained in simulation against real machine constraints, verified by deterministic gates before they touch hardware.
Trained and served in-house06 / Principles
How we build.
Sovereign by default
The serving path stays local. We train and run our own models on our own hardware, so the intelligence, and the data behind it, never leaves our control.
Embodied, not abstract
A model that only answers is unfinished. We measure ourselves on whether the machine moved correctly, on time, and safely, not on a benchmark score.
Verify by running
A confident wrong answer is worse than a crash. Every result is checked against ground truth by a deterministic gate before it is trusted.
05 / Infrastructure
GENOMA Pods: the compute we build ourselves.
Sovereign AI needs sovereign hardware. GENOMA Pods are L-series compute capsules: a printed double-lattice structure that carries its own thermal management and rack-level sensing, stacked with graphics cards. Capsules link into clusters, from a single rack to a full data hall, assembled on site without heavy equipment. More on GENOMA Pods →
Work with GENOMA Labs
Build the physical frontier with us.
Partnerships, pilots, and research collaboration on embodied AI, sovereign models, and autonomous heavy-equipment.