Five small models, one cheap shot at a frontier coder
An open-source workflow lets five modest Qwen models work as a team on hard coding problems. Each run costs under a dollar, and the post claims the team beats a frontier model.
Photo by Vishnu Mohanan on Unsplash
The claim, in the title of a Hacker News post: a team of five Qwen3.8-27B models beats a model called Fable 5 on LiveCodeBench Hard. The excerpt we have doesn't include the score table, so treat the head-to-head as the author's claim. The setup behind it is documented, and it is more interesting than the leaderboard line.
How the team works
The roles are a planner, two ideators, a solver and a manager, with a repair worker on call. No agent inherits another's chat history. Each starts clean, reads shared files and writes its output back to disk. That file trail is the only channel between them.
Success is not a model's opinion. A script runs the candidate program against sample inputs and outputs and compares the results. If it passes, the team is done.
The interesting part is the plumbing, not the models: files on disk, and a test script that decides who is right.
The method comes from a paper credited to Lee, 2026, known as GVS5H. The paper attributes the gain to the written-record discipline: every agent is told to work from the files, not from conversation.
The cost is the point
The Memdoor project rebuilt the authors' harness on the same model and ran the first ten problems from their LiveCodeBench Hard list, five passes each. That makes 50 runs, all graded on every hidden test. A run took a median of 40 minutes and $0.39. The whole batch cost $35.79.
Those are small sums for a benchmark that frontier labs treat as a showcase. The sample is also small: ten problems. One batch of 50 runs shows the approach is reproducible and cheap. It does not show it holds across the full benchmark.
What it signals
The workflow is eight YAML files and no code, and it names no model. Whatever model the user has pinned answers, including a local one. To change a role, you edit its prompt. That suggests the orchestration layer is becoming a commodity that anyone can swap models into.
Related tools point the same way. A Product Hunt listing, agentmash, pitches itself as open-source multiplayer AI for teams of coding agents. One listing proves little, but it shows builders reaching for coordination, not just bigger single models.
Our read, and it is speculation: if teams of small open-weight models keep closing the gap, the question for coding work shifts from which model is best to how well the process around it is built. In this design, the checks are plain tests and the memory is plain files, so the process is easy to inspect, cheap to rerun and runnable on your own hardware.