A local model and a cloud model can share a task instead of treating execution as an all-or-nothing choice. HazyResearch’s Minion and Minions work explored that design: local computation contributes to a larger reasoning workflow coordinated with a more capable model.
The original experiment ran the project’s Streamlit application and notebook using Ollama alongside a cloud provider. It compared the two protocols on a document containing hotel information and reservation dialogues.
Follow the information exchanged
The useful question is what each participant sees and produces. A model running locally does not make the combined workflow private if prompts, excerpts, intermediate results, or documents are sent to a cloud service.
The HazyResearch repository describes the protocols and their implementation. Minion uses interaction between the models; Minions adds task decomposition and aggregation across local work. Inspect the version you run rather than assuming every variant sends the same information.
Use a question with a known answer
The original task asked for a minibar bill from a longer hotel dialogue. This made it possible to compare the returned value against an identifiable passage instead of judging the answer only by its writing quality.
An interesting intermediate result was that one worker returned a room-rate figure while another found the relevant minibar charge. The aggregation step therefore had to distinguish a superficially relevant amount from the amount that actually answered the question.
That example illustrates why a final correct answer alone is not enough to understand the workflow. Inspect which pieces of evidence were extracted, what the workers misunderstood, and how the final model resolved the difference.
Measure the whole run
Record the model names, input document, protocol settings, requests, tokens, elapsed time, and final answer. The historical runs reported different cloud costs, but one small example is not a general savings estimate.
A stronger comparison includes a direct cloud baseline and questions of different difficulty. Extra coordination may be worthwhile for one workload and unnecessary for another. It also consumes local resources even when the provider bill falls.
Use the linked project for current setup instructions. This edition preserves the experiment’s central lesson: examine the actual exchange of data and work before deciding whether hybrid execution improves your application.