Stop Using Ollama

The charge sheet

A detailed post on Sleeping Robots alleges that Ollama rose to prominence by repackaging the work of others and then downplaying those origins. Zetaphor’s piece — which has been widely shared on Hacker News — traces Ollama’s early pitch as “Docker for LLMs,” its YC roots, and what the author frames as a long record of obfuscation about relying on Georgi Gerganov’s llama.cpp. It has been reported that Ollama’s documentation and binaries once omitted prominent credit and the MIT copyright notice required by llama.cpp’s license, and that community issues raising the problem went unanswered for months.

A technical pivot and a trust gap

Allegedly, Ollama later rebuilt its inference stack on top of ggml rather than continuing to use llama.cpp, arguing that stability and enterprise needs required a custom backend. The move, the post claims, reintroduced bugs and regressions that the llama.cpp community had already solved. It has been reported that the company also shifted toward cloud offerings and accepted venture funding — a pivot that, for critics, feels like abandoning the “local-first” ethos that drove early adoption. The emotional core here: users who trusted a tool to run models privately feel blindsided. Who wants their local LLM vendor to quietly flip the script?

So what now?

This isn’t just nitpicking. If the allegations are true, they speak to a larger question for the AI tooling ecosystem: how much credit and transparency do foundations of the stack deserve? Open-source projects thrive on attribution and reciprocity. The debate around Ollama is a reminder to inspect binaries, read licenses, and ask hard questions before betting your workflows on a single vendor. Alternatives exist, and the llama.cpp community is still very much alive — but the bigger takeaway is cultural: trust is hard to earn and easy to lose.

Sources: sleepingrobots.com, Hacker News