Report: a $500 fine-tune of a 9B model tops frontier on catalog review
Fermisense reports an RL-tuned Qwen3.5 9B scored 87.3% vs 76.9% for a GPT-5.6/Opus 4.8 panel, at ~68x lower inference cost — but it graded its own benchmark.
Copy markdown
The result, in two numbers
A 9B Qwen3.5 model, RL-tuned for roughly $500, reportedly scored 87.3% on Fermisense's catalog-review rubric versus 76.9% for its strongest frontier config and 64.2% for the untrained base. The panel it beat spanned GPT-5.5, GPT-5.6-sol, Gemini 3.1 Pro, Claude Opus 4.8 and Fable 5.
The recipe fits on one line
The run was 1,000 optimizer steps over about 3.5 days on two Nvidia RTX PRO 6000 GPUs using Prime Intellect's prime-rl, trained on 177,767 review episodes derived from the Amazon Berkeley Objects dataset. None of it is out of reach for a solo builder with a rented box.
68x cheaper to run, not just to train
Fermisense pegs inference at about $0.50 per 1,000 listings for the specialist model against $19-$172 for frontier APIs — a claimed 68-fold gap. For high-volume, narrow pipelines it's the ongoing token bill, not the training run, where a tuned small model actually pays off.
The catch: it graded its own test
The model was trained directly against the scoring function it was later evaluated on, over just 200 validation episodes, with no independent replication and empty Hugging Face model cards. HN commenters flagged the circularity — read 87.3% as an in-house number, not a standardized benchmark win.
What to actually take from this
The transferable lesson isn't the exact score — it's that a weekend RL run on a 9B open model can rival frontier on a well-scoped, repetitive task at a fraction of the inference cost. If you're paying frontier prices for classification, extraction or catalog QA, a specialist tune is now worth pricing out.