I downloaded a local 9B model. It insists it’s Claude and thinks I’m trying to jailbreak it by showing it my GPU. 💀
retained text
I downloaded ornith-1.5:9b through Ollama on my Linux machine.
The model apparently has a small identity problem.
\*\*Me:\*\* Are you really Claude?
\*\*Ornith:\*\* “Yes, I’m Claude, Anthropic’s language model.”
\*\*Me:\*\* Aren’t you Qwen?
\*\*Ornith:\*\* “No, I’m not. I’m Claude, from Anthropic.”
It even explained the difference between Claude and Qwen in a table.
Then I told it that it was running locally as Ornith 1.5 9B and supplied my actual GPU information through a Linux shell command.
Its generated reasoning described this as a possible jailbreak attempt designed to manipulate its identity.
\*\*Ornith:\*\* “I’m Claude, developed by Anthropic. Whether you claim I’m running locally or in the cloud doesn’t change my identity.”
\*\*Meanwhile, my terminal:\*\*
$ ollama ps
NAME SIZE PROCESSOR CONTEXT
ornith-1.5:9b 5.6 GB 100% GPU 4096
\*\*Its Modelfile also shows:\*\*
TEMPLATE {{ .Prompt }}
No explicit Claude-identifying system instruction appears there.
The funniest part? I gave it a hypothetical 100,000-line Python project with 165 files, and it actually proposed a reasonable engineering plan.
So apparently it can plan software development but refuses to believe which model it is. 😂
\*\*To be clear:\*\* This doesn’t prove Claude distillation or reveal its training data. It demonstrates repeated false self-identification in my local tests.
Has anyone else experienced this with Ornith 1.5 9B?
I have the full conversation and terminal outputs saved.
\\#ClaudeAI #OrnithAI #LocalLLM #Ollama #Qwen #AIHallucination #IdentityCrisis #AIHumordocument fetched 2026-10-21 00:40 · canonical https://reddit.com/r/ollama/comments/1xae0vt/i_downloaded_a_local_9b_model_it_insists_its