For a long time, the AI arms race has been about size. The logic was simple: bigger models equal smarter models. But a new experiment from Memdoor just flipped the script, proving that a coordinated team of 'smaller' models can absolutely dismantle a monolithic frontier giant.
The Power of the Pack
Memdoor didn't use one massive brain; they used a squad. By deploying five instances of the Qwen3.8-27B model, they created a specialized workforce. Instead of asking one AI to do everything, the workflow assigns specific roles: a planner, two ideators, a solver, and a manager, supported by a repair worker.
Interestingly, these agents don't chat in a fancy interface—they communicate strictly through files on disk. This lean, structured orchestration (based on the GVS5H method) allowed the team to tackle the 'Hard' problems on LiveCodeBench with surgical precision.
Crushing the Benchmark
The results are a wake-up call for the 'bigger is better' crowd. In tests on the first ten LiveCodeBench Hard problems, the Qwen team scored a staggering mean of 9.8 out of 10. To put that in perspective, Fable 5—a massive frontier model—managed an 8.0 alone.
What’s even more shocking is the cost. The entire operation, spanning 50 runs and 15.3 million output tokens, cost just $35.79. It turns out that specialized collaboration is not only more effective for complex coding but significantly more affordable than relying on a single, massive API call.
The Shift to Agentic Workflows
This isn't just a fluke; it's a signal. We are moving away from the era of the 'Omni-Model' and toward agentic ensembles. When you break a complex problem into planning, ideation, and verification, you reduce the hallucination rate and increase the rigor of the output.
If a team of 27B models can outclass the giants, the future of software engineering might not be one god-like AI, but a well-managed department of digital specialists.
Sources
Media
