Osmosis

Market position

A small, early-stage entrant in the model post-training and reinforcement fine-tuning infrastructure category, backed by Y Combinator (Winter 2025) and a $6.3M seed round reported in October 2025. Sources describe a team of roughly 2-10 people, listed as 6 on its YC profile.

Sources emphasize task- and domain-specific models trained with reinforcement learning that are claimed to beat general foundation models on accuracy, cost and latency; a hands-on, forward-deployed delivery model covering reward function design and serving; automated, continuous retraining without an engineer in the loop; and real-time learning from experience rather than reliance on static foundation models. Published benchmark claims include 21.5% first-run accuracy described as 60% better than base GPT-4o, and 7.5 average steps per task versus 67 for OpenAI Operator.

Technology

Reinforcement learning-based post-training of language models, including GRPO and DAPO algorithms and multi-turn tool training, with managed compute orchestration, reward modeling and training-run observability delivered through a CLI. The system integrates with customer evaluation tooling to trigger automatic retraining and can ingest real-time data to refresh models as often as hourly. Earlier work described a retrieval-based real-time learning loop using a vector database of past agent interactions with post-hoc performance scoring. The company has published Osmosis-Apply-1.7B, a reinforcement-learning fine-tuned code-merging model distributed via an MCP server, Ollama and Hugging Face, and reports it runs on local hardware.

Go-to-market

Direct sales via demo requests on the company website, developer-facing distribution of the CLI product and of open model artifacts through Ollama, Hugging Face and an MCP server, plus visibility through the Y Combinator network and launch posts. Founders solicited inbound contact directly by email in the YC launch.

Companies building and deploying AI agents and AI products, described as fast-growing AI companies and organizations in AI research and software development that need task-specific models, reliable multi-step tool use, document data extraction, or code generation. Specific customer names are not disclosed.

Geography

Headquartered in San Francisco, California, United States; job postings are also based in San Francisco. No other locations are disclosed.

History

Founded in 2024, the company participated in Y Combinator's Winter 2025 batch with Brad Flora as primary partner, launching an agent real-time learning framework benchmarked against OpenAI Operator. It subsequently released Osmosis-Apply-1.7B, a reinforcement-learning fine-tuned code merging model, and repositioned its public messaging around a broader "forward deployed reinforcement learning" post-training platform. A $6.3M seed round was reported in October 2025.

Risks & controversies

Public information is thin: the seed round source names no investors, no customers or revenue figures are disclosed, and performance claims (e.g., 21.5% first-run accuracy, 10X-50X faster training, cost multiples versus GPT-4o) come from the company's own launch and directory materials rather than independent evaluation. The company also operates under a different legal name (Gulp AI Inc.) and shares its brand name with unrelated entities, including Elsevier's Osmosis medical education product, the OsmosisAI visual inspection platform, and the Osmosis blockchain protocol, which creates naming-confusion risk.