TL;DR
- Fermisense fine-tuned Qwen3.5, a 9-billion-parameter open model, for about $500 in GPU time over three and a half days, and it beat GPT-5.6 Sol, Gemini 3.1 Pro, Claude Opus 4.8 and Claude Fable 5 on a catalogue-review task.
- The trained 9B model scored 87.3 per cent of the achievable score against 76.9 per cent for the best frontier configuration, at roughly 50 cents per 1,000 listings reviewed against $19 to $34.
- The same logic applies without any training: run a local model through Ollama for high-volume work with no judgement in it, such as reformatting text, extracting structured fields, or classifying into predefined buckets.
- Hold the exact figures loosely, since the result comes from Fermisense’s own benchmark on its own task and has not been independently reproduced.
In plain English
Fine-tuning takes an existing model and trains it further on one narrow task, which is how a 9-billion-parameter model can outperform much larger general ones inside that task’s boundaries. Parameters are the adjustable values a model learns during training and a rough proxy for its size. A frontier model is a current top-end system from a major lab, priced per token, meaning per chunk of text going in and out, which is where the cost comparison per 1,000 listings comes from.
The working pattern is model routing: choosing which model handles which step. A frontier model writes the instructions, the plan for what needs doing and how to check the work, then a local model running on the machine itself grinds through the volume. Ollama is the software that runs those models locally. This is slower, since an M2 MacBook Air cannot match an API call to a lab’s cluster, and the trade is worth it when the task needs little intelligence to begin with.
Fermisense published a case study on the 27th that’s been sitting near the top of Hacker News since: they took a 9-billion-parameter open model, Qwen3.5, trained it for about $500 in GPU time over three and a half days, and it beat every frontier model they tested — GPT-5.6 Sol, Gemini 3.1 Pro, Claude Opus 4.8, Claude Fable 5 — on a catalogue-review task. Not by a little. The best frontier configuration hit 76.9% of the achievable score. The trained 9B model hit 87.3%. Cost per 1,000 listings reviewed: about 50 cents for the specialist against $19 to $34 for the frontier options (AI Pricing Guru, Developers Digest). It’s their own benchmark, on their own task, not independently reproduced, so I’d hold the exact numbers loosely. But the shape of the result matches something I’ve been doing on a MacBook Air for months, at a much smaller scale, without ever writing it down as a strategy.
I run Ollama locally for what I’ve started calling the donkey work — the parts of a job that don’t need judgement, just volume. Reformatting a pile of text. Pulling structured fields out of something messy. Classifying a batch of anything into buckets I’ve already defined. None of that needs Opus or Codex sitting there thinking about it. What it needs is something cheap that will grind through five hundred of the same small decision without me paying frontier prices for each one.
The way it actually works day to day: I get Claude or Codex to write the instructions — the actual plan for what needs doing and how to check the work — and then hand execution down to the local model, which does the grinding while the expensive one stays free to think about the next problem. It’s slower. An M2 Air running a local model is never going to feel snappy next to an API call to a frontier lab’s cluster. But slower and correct, for free, beats fast and billed by the token when the task itself doesn’t require much intelligence to begin with.
Fermisense spent $500 and three and a half days proving that a small model, trained on the right narrow thing, can outright beat the frontier at its own game. I haven’t trained anything — I’m just routing. But the underlying bet is the same one: stop asking the most expensive model in the building to do work that doesn’t need it, and save that horsepower for the one decision in the pipeline that actually does.