A 1.1GB open-source small model cut fabricated answers from 10% to 0% across 131 fixed test questions. The cost: it searches only once, leaving it with just 24 correct answers out of 120 on live web pages. We believe this deserves more attention from enterprise IT teams than the usual pursuit of “larger models answering more questions.” Model capability depends not only on scale, but also on whether the boundaries of its task are clearly defined.
What this is
Obsessive-1.7B is a fine-tuned version of LFM3-1.7B that operates as a search subagent beneath a large model: an auxiliary model responsible for specific tasks such as search. After receiving a factual question, it makes exactly one web-search call, then returns a concise answer, a verbatim quotation, and a source link. If the search results do not contain the answer, it responds, “Not found.”
The developer trained it on roughly 9,000 tool-calling conversations and completed training on a consumer-grade laptop equipped with an RTX 4060. Fixed-result tests showed that it could find answers for 94% of questions that previously had answers and refuse every question without one. With live Wikipedia searches, however, it answered only 24 out of 120 questions correctly. It completed all 40 multi-fact questions, but fabricated information three times. The model also has limitations: it handles only single-turn conversations, works only in English, and occasionally cites adjacent facts.
Industry view
This project proves a simple judgment that is often overlooked: by narrowing a specialized small model’s responsibilities, teams can gain higher verifiability and lower resource costs. It does not need broad general-purpose intelligence to handle narrow jobs such as finding evidence, reporting sources, and admitting when it does not know. That makes it suitable for embedding in enterprise search, fact verification, and automated workflows.
The objections, however, are equally clear. A one-search limit causes the model to miss substantial amounts of available information, and its accuracy in live testing falls far below that of the original model. Its training data consists mainly of Wikipedia and FineWeb, so it does not represent complex commercial websites, paid databases, or Chinese-language use cases. More importantly, these are the developer’s own tests, with no independent reproduction. A fixed set of 131 questions is not enough to show that the model can operate reliably in production.
Impact on regular people
For enterprise IT: When budgets are limited, deploying a small model for narrowly defined tasks such as retrieval and verification may be more economical than purchasing a larger general-purpose model. Before deployment, teams must still test coverage and false-positive rates against real business materials.
For working professionals: Tools like this point toward professional AI assistants taking on more work involving finding materials, checking citations, and organizing evidence. Human value will increasingly shift toward asking good questions, judging whether information is reliable, and taking responsibility for the final conclusion.
For the consumer market: “Fewer answers, but clearer sources” is emerging as a new product selling point. The competitive focus of AI search may no longer be limited to how quickly answers are generated, but whether users can see where a conclusion came from and which parts genuinely could not be found.