GMAS claims 7B LoRA tops stock 14B

The interesting part here is not just the 7B-versus-14B claim, but the architectural bet that per-agent specialization can beat brute-force scale on consumer hardware. It is also still an unreleased, self-reported result, so the real test is whether the repo and benchmarks hold up once others can reproduce them.

• –Per-agent LoRA hot-swapping is a clever way to turn one smaller base model into a roster of specialists without keeping multiple full models in memory
• –Running this locally on M4 Max makes the project notable for edge inference and serious on-device experimentation, not just cloud demos
• –Custom Rust kernels suggest the team is optimizing beyond prompt orchestration and treating multi-agent systems as a systems problem
• –If the published benchmarks are real, this is a strong argument for smarter routing and specialization over simply moving to larger checkpoints