A $500 RL fine-tune of a 9B open model beat frontier models on catalog review

The point that the major labs don’t seem to get is that the vast majority of use cases simply don’t need models that have 50 PhDs and can speak 12 languages. Most use cases are defined within constraints where costs matter a lot.

As open weight models and cheap fine tuning services become the norm the whole economic framework of these mega models the labs are in an arms race building just completely crumbles. As does the economic picture that justified the massive infrastructure building that’s now broadly funded by a complex network of debt.

This is what makes open weight models so threatening to them. The political and “it’s China” angle is mostly just a cover for the real reasons why they’re freaked out.

The fact that models are now a pure commodity is bad enough for the big labs. If small open weight models become the norm the big labs are toast.

Every time I see this kind of story, two things bother me.

First, I have watched the free improvement of frontier models surpass the gains from retraining, many times now. Squeezing more out of the models that already exist, or simply doing nothing and waiting, is a real strategy and it often pays better. The fair comparison is not against today's frontier but against whatever ships while you are still maintaining your fine-tune.

Second, the $500 training bill is the cheapest line item in this story. The expensive parts are creating the data and maintaining the model afterwards. How many use cases can actually produce 177k scored episodes? Here they had to generate them synthetically from Amazon Berkeley Objects. To me, that dataset is the strongest evidence in the article of how hard fine-tuning is to apply: if the data existed naturally, nobody would need to manufacture it.

What I really started to notice is that SOTA models are really good at putting themselves out of the job.

We can see this already with GPT how luna can do 90% of what sol is used for. The only reason why china still bothers 'distilling' models is accurate training data generation, something that oai and anthropic had to spend years collecting while trying to dodge legal challenges.

The more intelligent models get, the more people will offramp to cheaper solutions that get the job done. There's no real benefit to using a sota model when the accuracy is already 99% and I think that is the biggest danger to US labs.

I have great interest in fine-tuning open models, and I'm looking for resources that HN folks can personally recommend. This article looks good and I've bookmarked it to read more thoroughly over time.

I've gotten as far as running Nemotron-3-Nano 30b locally, and plan to target models around 30b - 120b parameters. Based on brief examination of the results I can get, I think these vanilla models are capable enough to add real value, but training could push them over the finish line for specialized tasks.

What I really appreciate is that the author is thinking about the whole process, which is also my goal. Confirmation that others are identifying the same use case, and the same strategy for adding value using this technology.

This is a long term project, so I plan to buy hardware to conduct the fine-tune. ..

fine tuning a small LLMs or even real small language models (like bert) is what the recommended way since the introduction of LLM. the benefit is pretty much obvious: faster to run, fully controlling the stack, better fit for the custom domain...

but in practice not many people do the fine tuning, for pretty much a single reason: large language models API cost are still very cheap, fast enough, and get improvement all the times. what the point of spend time (and money) to fine tune a specified model, then just when you release it a newer gen generic model is released and beat it?

but if someday the progress for LLM is slowed, or the price increased to the point calling api is not a viable approach any more, then surely the day of fine tuning and small models will come again.

A fine-tune will generally outperform just about any other method on a closed domain, non generative problem.

LLMs are great because they can handle open domain problems in part because they are generative.

Amazing! I've been in a similar autoresearch-y rabbit-hole lately with getting Apple's 3B Foundation Model to match Sonnet 4.6 on a very specific task.

The result was: 90% parity achieved with a weird combination of a fine-tuned adapter + 1 deterministic step.

I wrote the whole thing down here: https://alexisrondeau.me/tada/research/FMDiscovery/dashboard... which includes the question, the answer, the 96 experiments and their lineage etc. etc.

I didn't read the Ramp article but this reads like a post hoc fallacy. Companies with 2x revenue have money to spend on AI. Companies with 1.15x revenue don't.

I like the 2x2 grid that describes when to fine-tune a model, when to use a frontier model, etc.

From the article it’s not clear how the scorer grades every episode - was it a frontier model that assigned the grade? How does that continue to work as the model that is being fine-tuned becomes better at the task than the frontier model?

I want small models to win and I am constantly experimenting with them. I have never tried fine-tuning and do not have that kind of budget. My approach is to remove some of the burden from models and bring into the agent.

Tool calling is an example - in some tasks RAG works really well, including coding agents where code, git log, Epics/Tasks, dependencies sources, etc. are all available in very structured manner. You can save many extra tool calls if you can run separate prompts and retrieve the source data needed for the actual work - rather its prompt.

And I really want to focus on search - this is the key technology if we want to use RAG instead of fine-tuning. If we can present really contextual sources in the prompts using a hybrid search approach - you can see how easily we get better results - either decisions or summaries from even small models.

turns out the bottleneck was never model size, it was having someone who actually understood the problem define the reward function

This continuous cycle of fine-tuned open models beating frontier on (often vaguely labeled/defined) benchmarks doesn't provide an accurate comparison to the expanding generalized capabilities of the SoTA, which makes them effectively meaningless.

If we were to take these at face value, why is it that the frontier labs' models are making legitimate new discoveries (e.g. Erdős and Jacobian conjectures) and these models are not?

To me, a better signal of capability would be similarly performing novel work at the same or better level, which they presently are not. I say this as someone who very much looks forward to open models being more capable, but to deny the gap is misguided hopeful hype.

"87.3%

Share of the maximum achievable score our GRPO-trained 9B open-source model reached on catalog review, vs 76.9% for the best frontier configuration: a 13.5% relative improvement over the frontier, and 36% over its own untrained base (64.2%). The five frontier models, even with optimized prompts, plateaued within a tenth of a point of each other; the trained specialist cleared that ceiling."

_______

This is hard for me to believe. I have a lot of skepticism that frontier models like GPT 5.5 that are likely 2T+ parameters in size only got about 12% more accurate than an untrained 9b parameter LLM.

The problem i've had with finetuning models is that most of the time better prompting beats finetuning

There seems to be no hold out data for test, so this is just overfitting?