tl;dr: frontier models are for renting; a model trained on your data is yours. these 10 startups are building the own-your-model stack: six platforms that turn your data into a model you own (Hyde, Thirdbrain Labs, Castform, Commissioned, distil labs, Conscious Engines) and four building the judgment layer that trains and proves it (Tacit, Physera, Vibrant Labs, Benchflow). the list was sourced with one sentence in claude code connected to frontrun's MCP: thesis in, companies + founders + funding out.
every business runs on rented intelligence today. the bet behind all ten of these companies is the same: the durable position is a model trained on YOUR data, YOUR judgment, YOUR workflows, that you own like you own your codebase. here's who's building that stack, and how this list was made in one prompt.
1. Hyde: training and inference for models enterprises own
Hyde is a training + inference platform for enterprises building specialist models they own. their line: "frontier models are for renting. your intelligence isn't."
founder: Anirudh Sanga (ex Palantir).
Frontier models are for renting. Your intelligence isn't. It's been incredibly validating to watch our core thesis play out across the industry. A massive shift is happening right now: companies are realizing that prototyping on giant, general models is...
— Hyde (@BuildwithHyde) July 10, 2026
2. Thirdbrain Labs: the post-training layer for models you own
Thirdbrain Labs is "the post-training layer that turns your data and expertise into models you actually own. data in. your model out." the team joined a16z speedrun on the thesis that every company becomes a portfolio of AI brains.
founders: Margaret Zhang and @latentius.
Every company will become a portfolio of AI brains. Almost none know how to build one. @latentius and I joined @a16z @speedrun to build @ThirdbrainLabs -- the post-training layer that turns your data and expertise into models you actually own. Data in. Your model out.
— Margaret Zhang (@_margaretzhang) April 13, 2026
3. Castform: RL post-training as a service for open-weight models
Castform is reinforcement-learning post-training as a service for open-weight models: bring your reward signal, they run the tuning and eval loop. their launch argument: "don't train your own model" is common AI advice, and your token bill is the proof it's wrong.
founder: Girish.
"don't train your own model" is common ai advice. it's wrong. your token bill's the proof. today, we're excited to launch castform into open preview. castform is the easiest way for you to train your own model, on your own data.
— girish (@googrish) June 11, 2026
4. Commissioned: custom models for non-technical teams
Commissioned fine-tunes models for non-technical people and small teams: custom intelligence for small businesses. bring your data in any shape, they handle the rest.
founder: Rediat B. Shamsu.
5. distil labs: a custom small model from a prompt, via distillation
distil labs trains a custom small language model from a prompt via distillation: train an SLM from your production traces and deploy it as a drop-in LLM replacement in a day, cutting LLM costs by up to 80%.
founders: Jan Golebiowski (ex Amazon ML) and @selimskii.
Next week we're rolling out a major update for local SLM workflows: run, refine, and evaluate. Minimal setup, reproducible results, private by default. stay tuned
— distil labs (@distil_labs) January 10, 2026
6. Conscious Engines: small specialist models embedded in products
Conscious Engines builds small, specialized AI models embedded in the products that already do the work: small, instant, invisible. their thesis in one line: a 1B specialist beats a 500B generalist at the job it was built for.
founder: Kautuk Kundan.
7. Tacit: RL environments for human judgment
Tacit builds reinforcement-learning environments for domains where human judgment is required: AI built on how your best people actually think, sharper with every interaction.
founders: Alex Dong and Joe Cole.
8. Physera: high-fidelity RL environments and real-world simulations
Physera is a research and product lab building high-fidelity RL environments and multimodal real-world simulations, at the intersection of model efficiency and behavioural simulation.
founders: Himanshu and Ashwarya Maratha.
Today we introduce Physera, a research and product lab rethinking applied intelligence. We are working at the intersection of model efficiency and behavioural simulations while building environments that are multimodal.
— Physera (@PhyseraAI) May 4, 2026
9. Vibrant Labs: autoscaling RL environments and benchmarks
Vibrant Labs builds autoscaling RL environments and benchmarks, from the team behind the ragas eval framework. their first open-source release: Enterprise Worlds, executable environments for training and evaluating agents on realistic enterprise workflows, starting with ITSMBench.
founders: ikka, @kranirudha, and @real_jjmachan.
Today, we're open-sourcing Enterprise Worlds: executable environments for training and evaluating AI agents on realistic enterprise workflows. First release: ITSMBench, an IT service management benchmark built around multi-turn tasks.
— ikka (@Shahules786) July 29, 2026
10. Benchflow: evals and environments for frontier capabilities
Benchflow builds evals and environments for frontier capabilities: the proving ground for what a model can actually do.
founder: Xiangyi Li.
a benchflow is born
— Xiangyi Li (@xdotli) May 1, 2026
how this list was actually made
one sentence in claude code, connected to frontrun's MCP. the agent took the thesis, ran a catalog keyword search plus a semantic thesis search over 42,000+ companies frontrun tracks from investor follow graphs, merged the results, resolved each company's founders, and checked for funding.
what used to be a week of manual sourcing was one prompt.
how to run this search yourself
frontrun's API and MCP are live for Pro users. connect it to claude code, cursor, or any MCP-capable agent and ask in plain english:
"find me companies building [your thesis], with founders and funding."
the same engine flags companies the day tracked investors converge on them: every round we flagged before it was announced, with dates and sources.
