A $500 Fine-Tune Beat Frontier Models — On One Simulated Task

Fermisense reports that reinforcement fine-tuning a 9-billion-parameter open model with GRPO — about 1,000 steps over three and a half days on two RTX PRO 6000 GPUs, roughly $500 of GPU time — produced a model that beat five frontier configurations on a catalog-review workflow, at $0.50 per 1,000 listings and around 68× lower per-unit cost than the strongest frontier baseline. Read the caveats before the headline: the workflow was simulated rather than production traffic, the evaluation has not been independently replicated, and the company selling this capability published it. The detail that should raise an eyebrow is one they report themselves — the model crossed into the frontier band after roughly a day of training, which is the classic signature of a reward function leaking task structure. What does generalize is the cost structure: producing a narrow-task specialist now takes two workstation GPUs and a long weekend, not a cluster.