Pick Osmosis when one narrow workflow, such as document extraction or tool-calling agents, costs too much on a frontier API and you want the trained weights back. Expect a hands-on engagement with no published price, a small team, and a trust center showing SOC 2 Type 2, HIPAA and GDPR as compliant.
Osmosis, legally Gulp AI Inc., is a San Francisco reinforcement fine-tuning company from Y Combinator's Winter 2025 batch, founded in 2024. It trains task-specific models with GRPO and DAPO, publishes open-weight models on Hugging Face, and lets customers export the weights they train.
Founded: 2024 · HQ: San Francisco, CA, USA · Team: 1-10 · CEO: Kasey Zhang · Funding: Y Combinator (Winter 2025) plus operator angels named on its About page; round sizes not confirmed
About Osmosis
Osmosis is a post-training company incorporated as Gulp AI Inc. and based in San Francisco. Its About page says it was founded in 2024 by Kasey Zhang and Andy Lyu, friends since middle school, and it went through Y Combinator's Winter 2025 batch. Zhang is the CEO and a repeat founder who sold his first company as an undergraduate. Lyu is the CTO and was the youngest tech lead on the real-time recommendations and model serving team at TikTok.
The product is a reinforcement fine-tuning platform the company calls forward deployed reinforcement learning. The docs describe the split plainly: you define or port your agent loop, tools, rewards and training data, and Osmosis runs the training and serving to produce task-specific models that it says can outperform foundation models on performance, cost and latency. It names GRPO and DAPO as the algorithms, supports multi-turn tool training with tools synced from GitHub or hooked in over MCP, and can retrain on a schedule or an evaluation trigger as often as every hour. The three use cases it puts forward are document data extraction, agents that use their production toolkit reliably, and small coding models for domain-specific languages and front-end components. A Python SDK and CLI, published under the MIT licence on the company's GitHub organisation and updated in late October 2026, is how developers and agents drive it.
Ownership is the clearest selling point. The platform page says you own the weights you train and can serve them on Osmosis single-tenant compute or export them to self-host, and it says the whole platform can be deployed inside your own cloud or on-prem to keep data in your boundary. That puts it in a different position from the managed fine-tuning offered by OpenAI, where the tuned model stays inside one provider, and nearer to hosted trainers such as OpenPipe and Together AI, or inference hosts such as Fireworks AI and Replicate, that also sell tuning.
It also publishes open weights, which is unusual for a firm whose business is the training service. Three models sit under the osmosis-ai organisation on Hugging Face, all Apache 2.0: Osmosis-Structure-0.6B (about 416 likes, the most liked), Osmosis-Apply-1.7B (access-gated) and osmosis-mcp-4b. They date from mid-2025 and serve as public demonstrations of the method. The engineering blog is more recent: its latest posts, from April to June 2026, cover LoRA training for mixture-of-experts models and cutting memory in long-context reinforcement learning.
Trust and compliance are stated on a vendor-run trust center, which marks the company compliant with SOC 2 Type 2, HIPAA and GDPR, next to a long controls list that includes a penetration test within the last twelve months. Those are the company's own statements, so a buyer handling regulated data should still ask for the audit report. Investors the About page names are Y Combinator and a list of operator angels including Paul Graham, Guillermo Rauch of Vercel, Misha Laskin of Reflection AI and Erik Bernhardsson of Modal; the size of its funding rounds is not confirmed on any page we could read.
The other side of the argument is capacity and price. There is no published pricing and every call to action is a demo booking, so you cannot size a project before a sales conversation. A hands-on service from a small team limits how many engagements it can run at once, and a team that has already shipped a reinforcement fine-tune can reach the same published algorithms with open-source trainers and rented GPUs, trading the vendor's reward design for lower cost and control. For a broader view of the cost trade-off against hosted APIs, see our guides on choosing the right LLM and the cheapest LLM API in 2026. The comparison favours Osmosis when the job is narrow, the volume is high and ownership of the weights matters.
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/platform
- Osmosis SDK and CLI (Python SDK and CLI (MIT)): https://github.com/Osmosis-AI/osmosis-sdk-python
- 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
Compliance
SOC 2 Type II, HIPAA, GDPR
Links
Frequently Asked Questions
Does Osmosis publish a price?
No. The site shows no plans and sends visitors to a demo request, so the cost depends on the workflow being trained and how much hands-on work it needs. Ask for a scoped quote and for what the training compute is billed as before you compare it with a per-token API.
Who is behind Osmosis and who has backed it?
Kasey Zhang is the CEO and Andy Lyu the CTO. It is a Y Combinator company, and the About page lists operator angels including Paul Graham, Guillermo Rauch, Misha Laskin and Erik Bernhardsson. No round size is public on the pages we found.
Which models has Osmosis released?
Three, each released as open weights with an Apache 2.0 licence: Osmosis-Structure-0.6B handles structured output, osmosis-mcp-4b handles tool use, and Osmosis-Apply-1.7B sits behind an access gate. They are about a year old and work as demonstrations of the training method.
Can I keep the model Osmosis trains for me?
Yes, according to the vendor. Its platform page says trained models belong to you, to be served on dedicated single-tenant servers or taken away for self-hosting, and that the whole platform can run in your own cloud or on-prem. Confirm the export format and licence terms in the contract.
Is Osmosis suitable for regulated data?
The trust center shows SOC 2 Type 2, HIPAA and GDPR compliance, and the VPC option keeps training data inside your boundary. These are the vendor's own statements, so request the audit report and a signed data processing agreement, which the site publishes as a template, before sending sensitive records.