Together AI suits developer and research teams that want to train or serve open-source models instead of depending on closed APIs from proprietary providers. Backed by over $50 million from venture investors and a San Francisco team of 50 to 100, it trades some of that polish for lower infrastructure costs and full control over model weights.
Founded in 2022, Together AI is a San Francisco cloud infrastructure company that lets researchers and organizations train and serve open-source language models on distributed GPU clusters instead of building that infrastructure themselves. Its RedPajama dataset and OpenChatKit models are among its notable open releases.
Founded: 2022 · HQ: San Francisco, California, USA · Team: 50-100 · CEO: Isaac Ong · Funding: $50+ million · Valuation: $600+ million (estimated)
About Together AI
Together AI was founded in 2022 by Isaac Ong, Tim Tully, Ramesh Chandra, and others from AI research institutions and technology companies. The company is headquartered in San Francisco with roughly 50 to 100 employees, has raised more than $50 million in venture funding, and carries an estimated valuation of $600 million or more. Its platform lets researchers and organizations train large language models across distributed GPU clusters without owning that hardware themselves; Together Compute handles the multi-machine coordination behind that training, and the Inference Engine then serves the resulting models at scale. The API Service exposes hosted endpoints for models trained on the platform, and integrations with PyTorch and Hugging Face let teams bring existing training code over with minimal changes. The company co-led RedPajama, an open dataset and model project that reproduces large training datasets for language models, and released OpenChatKit as one of the open models built to demonstrate the platform. Together publishes research on training efficiency and distributed systems and maintains partnerships with research institutions working on open-source AI. Together AI positions itself against providers that keep training infrastructure and model weights proprietary, arguing that wider access to compute lets more organizations compete in model development. The company documents model cards for the open-source models it trains, covering architecture, training data composition, and identified limitations.
Mission
To democratize access to AI infrastructure enabling diverse organizations and researchers to train and deploy language models efficiently and cost-effectively.
Products
- Together Compute
- Inference Engine
- API Service
- Integration Ecosystems
Compliance
SOC 2 Type II
Links
Website · GitHub · Twitter · LinkedIn · Blog · Docs
Frequently Asked Questions
How much funding has Together AI raised, and what is it worth?
Together AI's valuation is estimated at $600 million or more, funded by venture investors who see distributed training and inference infrastructure as strategically important. The company has used that funding to expand its GPU compute capacity and grow its engineering and research teams out of San Francisco.
What products does Together AI make, and how is pricing structured?
Together AI's product lineup, Together Compute, the Inference Engine, and the API Service, is billed per token for serverless inference and by the hour for dedicated GPU instances. Exact rates vary by model and hardware and are published on Together's pricing page rather than fixed company-wide.
Is Together AI SOC 2 compliant?
Yes, Together AI holds SOC 2 Type II certification. It operates infrastructure across US, EU, and Asia-Pacific data regions, and publishes its privacy policy and a trust center covering platform reliability.
How does Together AI compare to Groq in 2026?
Together AI runs model training and inference on distributed GPU clusters that customers can also use to fine-tune their own models, while Groq built a custom LPU chip purely for low-latency inference and has licensed that chip design to NVIDIA. Teams that need to train or customize open models pick Together; teams that only need fast inference on models that already exist often look at Groq instead.
How do you get started with Together AI?
Developers can sign up for the API Service to call hosted open-source models directly, with documentation and quickstart guides at docs.together.ai. Teams that want to train their own models submit training code and datasets to Together Compute for distributed GPU training instead.
Top Alternatives
- Groq: Pick Together for training plus inference on your own infrastructure; pick Groq for turnkey low-latency inference on its LPU chips.
- SambaNova Systems: Pick Together for open GPU-based training and inference; pick SambaNova for its custom RDU chips built specifically for inference.