A 1.1 GB operational contract

Obsessive-1.7B does not chat. The model, derived from LFM3-1.7B and about 1.1 GB, performs a single task: it takes a factual question, calls web_search once and returns a short answer, a word-for-word citation and the source URL. If the search results do not contain the answer, it replies "Not found in the search results." It is not designed for conversation; its place is below a larger model as a subagent producing verifiable items. This division of labor is not a technical footnote: it shifts how a local LLM is judged, from conversational ability to the capacity to maintain a strict operational contract.
The small footprint and GGUF deployment, with llama-server and a 16,384 token context, place Obsessive-1.7B in the self-hosted perimeter. On a laptop with an RTX 4060 the median time per question drops from 5.1 to 2.0 seconds compared with the base model. This makes the subagent a realistic component for on-premise pipelines, not a lab curiosity.
For teams evaluating local infrastructure, the signal is clear: strong specialization can produce, with a few gigabytes and a consumer GPU, behavior that is more predictable than a larger generalist. The question is no longer whether a small model is intelligent, but which role it is given and what guarantees it offers the orchestrator.

Coverage versus predictability: numbers at a fork

Fixed-benchmark tests (131 SimpleQA questions, 29 with the answer removed) expose the fork. The base model finds the answer when present in 50% of cases; Obsessive does so in 94%. When the answer is missing, the base never says "not found", while Obsessive does so in 100% of cases. Invented answers fall from 10% for the base to 0%. Operationally, a missing answer and a wrong answer have different costs: absence can be managed with a retry, fabrication requires downstream verification or produces damage.
On live Wikipedia search, with 120 questions and 40 multi-fact requests against a backend never seen during training, coverage flips. The base model returns a response in 38% of cases, Obsessive in 100%. On correct answers, however, the base wins: 37 out of 120 versus 24 out of 120. The reason is stated: the base keeps searching and writes long answers, so the answer emerges more often, but in 62% of cases it exhausts its turns and returns nothing. Obsessive searches once and declares only what it can cite.
The structural point emerges here. In an automated system, a component that does not respond in 62% of cases is a black hole: the orchestrator does not know whether to retry, move on or stop. A subagent that always returns an answer, with citation and URL, offers a tighter operational contract. It pays in raw coverage: it says "not found" for about one in ten questions where the answer is actually in the results and sometimes cites a nearby fact, such as the wrong year's winner. But the architecture allows downstream verification: the citation is there, ready for checking. On fabricated factual values, the live comparison is four out of four for the base and three out of fifty-two for Obsessive.

Laptop fine-tuning changes the geography of training

The training method deserves attention. Producing Obsessive-1.7B used LoRA rank 16, about 9,000 tool-calling conversations and an RTX 4060 laptop. No cluster is required: fine-tuning happens on consumer hardware. This lowers the threshold for creating specialized subagents and brings training back inside the organization.
The data come from Wikipedia and FineWeb pages. Each answer and citation was checked to appear verbatim; about a quarter of the data teaches "not found". This is not just positive examples: the dataset also trains refusal when the sources are insufficient. It is data engineering, not model magic.
For teams evaluating on-premise deployment, this combination changes constraints. Training data control does not require cloud infrastructure; GGUF distribution uses tools already adopted. Fine-tuning cost is tied to a consumer GPU and a curated dataset, not a recurring API bill. TCO shifts from cloud inference spending to maintaining local components and data quality.

Verifiable subagents: an accounting unit for pipelines

In a multi-model architecture, the orchestrator delegates factual tasks to subagents. Obsessive-1.7B's value is not knowing more, but always returning a verifiable object: answer, verbatim citation and URL. This stable shape lets the orchestrator apply rules, logs and checks without adapting to unpredictable formats.
The comparison with the base model is instructive. A model that sometimes does not respond, sometimes invents, forces upstream uncertainty management. A subagent that always responds, even with "not found", turns failure into a coded event. Operational predictability becomes a first-order metric, alongside answer quality.
Downstream verification is the approach's strength. The citation is already extracted and the link is ready: a control system can compare the quoted text with the source without regenerating or asking the model. For regulated sectors or internal archives, this reduces the risk of introducing unverified content into automated procedures.
There is a cost: raw coverage falls. Obsessive says "not found" in about one in ten cases where the answer was present and sometimes cites the wrong fact. But an error with a citation is cheaper to detect than a fluent answer with no source. The choice between a higher-coverage model and a more verifiable one is not strictly technical: it is a risk decision.

Sovereignty and costs: the return of the specialized model

The idea of a 1.1 GB local LLM specialized as a subagent shifts incentives for teams evaluating on-premise pipelines. No data center is needed for specialized components: fine-tuning can happen on a laptop, data stays in-house and deployment goes through known GGUF tooling. The comparison is no longer between a large generic model and a small useless one: it is between statistical coverage and operational predictability.
Sovereignty becomes concrete. An organization can fine-tune a subagent on its own documents, with its own citation rules, without sending queries to a cloud provider. Model and context data remain within the corporate perimeter. For regulated sectors this reduces legal and compliance risks.
On TCO, a local component eliminates per-token costs from external APIs and reduces network latency. The fixed cost is the hardware for inference and for any fine-tuning cycles. With an RTX 4060 laptop the median time is already compatible with many automated workflows; no datacenter GPU is required. Multiplying specialized subagents can fragment maintenance, but each component is small and replaceable.

What to watch: multi-fact, multi-turn and trust thresholds

Obsessive-1.7B's stated limits are consistent with its chosen constraint: one question per conversation, multi-turn confuses it, English only, and it tends to report claims as facts. It is not a general assistant substitute nor a search engine. It is a component with a narrow contract.
The next step is to understand whether the model can handle multi-fact questions without losing predictability. The 40 multi-fact requests in the live test suggest the problem is already on the table. A subagent that answers one question at a time forces the orchestrator to split the work and reassemble results, with coordination costs. The challenge is maintaining citation verification when the task becomes composite.
Multi-turn also needs observation. Today the model gets confused if the conversation continues. For pipelines that query documents iteratively, this is a strong limit. Specialized variants for short sessions or iterative querying may emerge, perhaps with different contracts.
Finally, trust thresholds deserve reflection. The model offers deterministic behavior with temperature 0, but citation quality is not uniform. Monitoring false "not found" rates and citation precision will become as important as measuring accuracy. For AI-Radar users, the analytical framework on /llm-onpremise is meant to weigh these trade-offs without turning an experiment into a universal recommendation.