Ollama Raises $65M Series B — 9M Developers Already Run Models Locally
Ollama announced a $65M Series B on July 9, led by Theory Ventures with Benchmark, 8VC, Y Combinator and Pace Capital participating. Total funding is now $88M. Valuation undisclosed — the founders also declined to talk revenue. The numbers they did share are the story: 8.9 million monthly active developers and more than 67,000 community-built integrations, with partnerships across every major model lab and hardware vendor.
By 2026 standards $65M is pocket change — frontier labs burn that on a single training run. But Ollama owns something no lab does: the default. When anyone wants to run an open model on their own machine — Qwen, GLM, Kimi, gpt-oss, Gemma — the first command they type is ollama pull. Every open-weights release from a Chinese or American lab makes that position more valuable, and the past two months have been a nonstop parade of them.
The agent angle is what makes this round interesting. Local inference used to be a hobbyist thing; now it is the cost-and-privacy half of the agent stack. Agents burn tokens continuously, so routing the routine 90 percent of calls to a local model is obvious economics — local-first memory systems have been trending on GitHub for exactly the same reason. The open question is monetization: 8.9 million users on a free tool, and the paid cloud tier now has to prove that the company built on local can also sell hosted. Theory and Benchmark just bet $65M that owning the developer default is worth it.
https://ollama.com
← Back to all articles
By 2026 standards $65M is pocket change — frontier labs burn that on a single training run. But Ollama owns something no lab does: the default. When anyone wants to run an open model on their own machine — Qwen, GLM, Kimi, gpt-oss, Gemma — the first command they type is ollama pull. Every open-weights release from a Chinese or American lab makes that position more valuable, and the past two months have been a nonstop parade of them.
The agent angle is what makes this round interesting. Local inference used to be a hobbyist thing; now it is the cost-and-privacy half of the agent stack. Agents burn tokens continuously, so routing the routine 90 percent of calls to a local model is obvious economics — local-first memory systems have been trending on GitHub for exactly the same reason. The open question is monetization: 8.9 million users on a free tool, and the paid cloud tier now has to prove that the company built on local can also sell hosted. Theory and Benchmark just bet $65M that owning the developer default is worth it.
https://ollama.com
Comments