# Osmosis

> Osmosis is a San Francisco reinforcement learning platform (YC W25) that helps companies fine-tune open-source models to beat foundation models on cost, latency, and reliability. Founded by Kasey Zhang and Andy Lyu, it lets teams train task-specific AI agents with multi-turn tool use over the Model Context Protocol, then export or self-host the weights they own.

- **Founded:** 2024
- **Headquarters:** San Francisco, California, United States
- **Founders:** Kasey Zhang (Co-Founder & CEO), Andy Lyu (Co-Founder & CTO)
- **Team size:** 12
- **Products:** Osmosis Platform, Multi-turn Tool Training over MCP, Deployment & Export, Osmosis Agent Toolkit
- **Notable:** Raised a $6.3M seed round co-led by CRV and Audacious Ventures in 2025., Backed by notable operators including Paul Graham, Guillermo Rauch, Misha Laskin, Erik Bernhardsson, and Shane Gu., Graduated from Y Combinator's Winter 2025 batch.

## Products & services

- **Osmosis Platform** — A reinforcement fine-tuning platform for training task-specific AI agents. Handles data processing, compute orchestration, reward modeling, and training-run observability, with real-time RL, automated hyperparameter sweeps, and support for training and maintaining hundreds of model versions.
- **Multi-turn Tool Training over MCP** — Turns on multi-turn tool training so agents learn to use tools reliably. Tools can be built natively or synced from GitHub, and hooked in over the Model Context Protocol to replicate production conditions.
- **Deployment & Export** — Serve trained models on single-tenant compute or export them for self-hosting - customers own their trained weights, with full VPC/on-prem deployment available.
- **Osmosis Agent Toolkit** — An open-source (MIT) collection of TypeScript packages that lets LLMs interact with the Osmosis protocol and tools.

## Achievements

- Raised a $6.3M seed round co-led by CRV and Audacious Ventures in 2025.
- Backed by notable operators including Paul Graham, Guillermo Rauch, Misha Laskin, Erik Bernhardsson, and Shane Gu.
- Graduated from Y Combinator's Winter 2025 batch.
- Published open-source releases including an SLM trained for MCP and a Qwen3-1.7B model fine-tuned for code merging.
- Released research on training thousands of LoRA adapters at once and cutting memory in long-context RL with fused logprobs.

## Latest updates

- **2026-06** — Published research on cutting memory in long-context RL using fused logprobs for multi-turn agentic applications.
- **2026-06** — Detailed training thousands of LoRA adapters at once by sharing base models and routing tokens.
- **2026-04** — Added LoRA MoE RL support for Qwen3.5 models using Megatron-LM and SGLang.
- **2025-10** — Launched publicly with a $6.3M seed round co-led by CRV and Audacious Ventures.

## Links

- Website: https://osmosis.ai
- LinkedIn: https://www.linkedin.com/company/osmosis-ai/
- Twitter/X: https://x.com/Osmosis_AI
- GitHub: https://github.com/Osmosis-AI

---

Profile page: https://yespress.io/osmosis-yc-w25
Published by YesPress — https://yespress.io
Last updated: 2026-07-30