Osmosis is worth knowing if you want one narrow workflow done cheaper than a frontier API can do it. Roughly six people, backed by CRV and angels who build infrastructure themselves, giving weights away while charging for the training service. Early, hands-on, and capacity-limited by design.
Osmosis is a San Francisco post-training company from Y Combinator's Winter 2025 batch, incorporated as Gulp AI Inc. It sells reinforcement fine-tuning that produces task-specific models, using published methods such as GRPO and DAPO, and gives its own model weights away openly on Hugging Face.
Founded: 2025 · HQ: San Francisco, CA, USA · Team: 1-10 · CEO: Kasey Zhang · Funding: About $7M total following a $6.3M round (Y Combinator, CRV, Audacious Ventures)
About Osmosis
Osmosis is a post-training company founded in 2025 and incorporated as Gulp AI Inc., based in San Francisco and part of Y Combinator's Winter 2025 batch. It was co-founded by Kasey Zhang, who is CEO and had previously co-founded and sold a gaming startup before working in early-stage AI venture investing, and Andy Lyu, the CTO, a real-time data and machine learning engineer who was the youngest tech lead at TikTok. The team is roughly six people. The product is a reinforcement fine-tuning platform the company describes as forward deployed reinforcement learning. Customers use it to train task-specific and domain-specific models intended to beat general foundation models on the narrow job they were trained for, measured on accuracy, cost and latency rather than broad capability. The platform handles compute orchestration, reward modelling and training-run observability, and is delivered as a CLI product usable by both developers and agents. Technically it leans on published reinforcement learning methods rather than a proprietary black box, naming GRPO and DAPO as the fine-tuning techniques, with a continuous-improvement loop that can retrain a model as often as hourly. The three use cases the company puts forward are domain-specific document extraction, training agents to use tools across multi-step workflows, and generating specialised coding models for in-house languages and component libraries. Osmosis publishes open weights, which is unusual for a company whose business is the training service rather than the model. Four models sit under the osmosis-ai organisation on Hugging Face, all Apache 2.0: Osmosis-Structure-0.6B, osmosis-mcp-4b, Osmosis-Apply-1.7B and an ONNX classifier head. Osmosis-Structure-0.6B is the most adopted, at roughly 8,000 downloads and over 400 likes, and doubles as the public demonstration of the method the paid platform sells. Funding is early. Public reporting puts the total at about $7M following a $6.3M round, with Y Combinator, CRV and Audacious Ventures as institutional backers. The angel list is notable for a company this size: Paul Graham, Erik Bernhardsson of Modal Labs, Misha Laskin of Reflection AI, and Guillermo Rauch of Vercel, which reads as infrastructure-operator conviction rather than generalist money. Competitively it sits between the managed fine-tuning offered by the large model providers and the do-it-yourself path of running verl or TRL on rented GPUs. Against the former it offers open weights and no lock-in to one model family; against the latter it offers the reward modelling and orchestration most teams get wrong. The obvious risk is scale: a six-person company selling a forward deployed service has a hard ceiling on how many customers it can serve at once, and no published pricing to signal where that ceiling sits.
Mission
Help companies create task-specific models that beat foundation models at a fraction of the cost.
Products
- Osmosis platform (Reinforcement fine-tuning platform): https://osmosis.ai/
- Osmosis-Structure-0.6B (Open-weight model): https://huggingface.co/osmosis-ai/Osmosis-Structure-0.6B
- osmosis-mcp-4b (Open-weight model): https://huggingface.co/osmosis-ai/osmosis-mcp-4b
- Osmosis-Apply-1.7B (Open-weight model (gated)): https://huggingface.co/osmosis-ai/Osmosis-Apply-1.7B
Osmosis Models on HokAI
Links
Frequently Asked Questions
How much does Osmosis cost in 2026?
Osmosis publishes no pricing. Engagements are scoped per customer as a forward deployed service, so the cost depends on the workflow being trained and the depth of hands-on work. Budget separately for the GPU compute the training runs consume, which is not covered by any published platform figure.
Who founded Osmosis and who backs it?
It was co-founded by Kasey Zhang, the CEO, who previously sold a gaming startup and then invested in early-stage AI, and Andy Lyu, the CTO, formerly the youngest tech lead at TikTok. Institutional backers are Y Combinator, CRV and Audacious Ventures, with angels including Paul Graham, Erik Bernhardsson, Misha Laskin and Guillermo Rauch.
What models has Osmosis released?
There are four, every one of them Apache 2.0 licensed. Osmosis-Structure-0.6B targets structured output generation and is the most adopted, osmosis-mcp-4b targets tool use, and Osmosis-Apply-1.7B is access-gated. Releasing weights while charging for the training service is the unusual part of the strategy.
What are the best alternatives to Osmosis?
Managed fine-tuning from the large model providers is the low-effort option if you are content staying inside one model family. Running verl or TRL yourself on rented GPUs is cheaper and gives full control, provided you can do your own reward modelling. Pick Osmosis when you want the reward design handled by people who do it repeatedly.
How does Osmosis compare to doing reinforcement fine-tuning in-house in 2026?
The methods Osmosis uses are public and the tooling is open source, so nothing stops a capable team doing it internally. What you are buying is the reward modelling judgement and the orchestration, which is where in-house attempts usually stall. If you already have an ML team that has shipped a reinforcement fine-tune, in-house is cheaper.