TypeSafe AI
Machine-native, composable AI: Jev, a 'System One' model priced per billion tokens, 100x cheaper than LLMs.
A NeoLab is an AI research company started by people who left DeepMind, Meta, OpenAI and xAI or a top university lab, usually valued at a billion dollars or more before it sells anything. We track 102 of them: what each is building, where it is, how much it has raised and from whom, and whether anyone outside can use what it makes. Everything refreshes every three days, and every number links to its source. See which labs came from where.
Funding, launches, deals and departures across the tracked labs. Newest first.
Machine-native, composable AI: Jev, a 'System One' model priced per billion tokens, 100x cheaper than LLMs.
Action and world models trained on Medal's two billion gameplay videos, for agents that need spatial and temporal reasoning.
An industrial AI model for physical industries: a self-improving operational-intelligence platform; incubated by Eclipse.
Reinforcement learning as a service: a cheaper, more open alternative for enterprises; ~$50M ARR.
Open-weight hybrid reasoning models built by iterated distillation and amplification; post-training for self-improvement.
Only valuations set by a disclosed round. Estimates from trackers are shown on the lab pages, marked as estimates.
Long-horizon labs aiming at AGI or superintelligence, usually with no near-term product.
Safe Superintelligence, Thinking Machines Lab
Systems that run the experiments: AI that does AI research, recursive self-improvement.
Ineffable Intelligence, Recursive Superintelligence, Mirendil, Core Automation…
Models that learn physics and 3D structure from video and interaction, for games, robots and simulation.
General Intuition, World Labs, Decart, AMI Labs…
Robot foundation models and general-purpose physical intelligence.
Skild AI, Physical Intelligence, Generalist AI, Dyna Robotics…
Autonomous labs, AI scientists, materials and molecular discovery.
Lila Sciences, Periodic Labs, CuspAI, Edison Scientific…
Formal proofs, autoformalisation and verified reasoning.
Harmonic, Axiom Math, Logical Intelligence, Math Inc…
Code models and agents built as research labs rather than dev-tool startups.
Poolside, Cognition, Magic
Models that act on screens and in software, trained from demonstrations.
Isara, Standard Intelligence, H Company, Prentis
Labs whose route to market is open weights and open training stacks.
Reflection AI, Applied Compute, Nous Research, Essential AI…
Diffusion LLMs, state-space models, category theory, continual learning, data efficiency.
TypeSafe AI, Sakana AI, Liquid AI, Fundamental…
Labs whose thesis is the substrate: chip design, neuromorphic hardware, kernels.
Unconventional AI, Ricursive Intelligence, Inferact, RadixArk…
Models designed around people, relationships and personal context.
Hark, Humans&, River AI
Generative voice, image and video labs with a research-first posture.
Luma AI, Black Forest Labs, Cartesia, Kyutai
Protein design, molecular structure and AI-first drug discovery.
Chai Discovery, Merge Labs, EvolutionaryScale, Isomorphic Labs…
Mechanistic interpretability, oversight and non-agentic AI, often as a public-benefit company or nonprofit.
Goodfire, Irregular, LawZero, Transluce