A heating engineer built a swarm of six small LLMs: splitting work into sections gave the weak model a ×4 gain

A department head at a design organization, a heating engineer with fourteen years in design work, set out this spring to grow a swarm of weak local models that would outperform the strongest model in the group. He has never been a programmer, and he has worked with neural networks for less than a year. The swarm did not get smarter. The measurements did show, however, that of the whole design-institute org structure, only one element carries the load: splitting the task into sections. It gave the weak model roughly a fourfold gain, while the swarm as a whole, judging by the author's headline, added about five points.

A caveat: the source is only the announcement of the post on Habr, with no tables or methodology. What follows covers what the announcement claims, and the figures and their units should be checked against the full text.

An institute instead of a pack

Anyone who has worked in a design organization will recognize the idea. No single person solves a task alone. There is a project lead, departments responsible for sections, reviewers, and sign-offs. The author arranged six small models along these lines and checked which elements produce results and which only add correspondence between agents.

The hypothesis was typical of multi-agent experiments: if you gather many weak executors and put a management layer above them, the combined quality will exceed that of the best individual member. That hypothesis did not hold up. The swarm did not beat the strongest model in its own lineup.

What the measurement showed

The negative result is more useful here than a positive one would be. The author did not simply report that the swarm failed to take off. He isolated one mechanism that works. The main points of the announcement:

• Setup: six small local models of varying strength.
• Goal: a swarm that outperforms the strongest model in the set. Not achieved.
• Splitting the task into sections: roughly a ×4 gain for the weak model.
• The swarm as a whole: about five points, per the headline. The announcement does not disclose the units of measurement.
• Asymmetry: the weak model gains more from the split than the strong one.

A plausible explanation, which is our interpretation and not a quote from the author: a small model struggles to hold a long prompt with several requirements at once. When it is handed one narrow section with a clear input and output, it handles a task it previously fell apart on. A strong model has more context and reasoning headroom, so decomposition helps it less. The institute-style "reviewer" role and the other layers of sign-off apparently added no gain, or far less.

What to do with this in practice

For those building local pipelines on small models, the takeaway is down to earth. Before constructing a hierarchy of lead, executors, and reviewers, check one thing first: what plain decomposition into independent parts, with rigid input and output templates, gives you. It is cheaper in tokens and latency than multi-layer orchestration, and, judging by the author's measurements, it is what delivers most of the effect.

Second, the baseline should not be the swarm's average model but the strongest one available. A swarm that does not outperform its best member is, in effect, paying for extra calls with nothing in return.

Third, there is the format itself. It is rare for an experiment to come from someone in a neighboring industry, bringing an engineer's view of how work is organized. Design institutes have spent centuries working out how to divide responsibility without losing quality. Which parts of that experience carry over to LLMs, and which do not, is best judged from the author's full text, which should include the methodology, the task set, and a breakdown by model.