Interviews · Interview clip
When to fine-tune an open model instead of using a frontier model
Cursor cut cost per task by 86% by moving a specific workload away from a frontier model and fine-tuning a smaller model for that job.
That is where open-weight models can become very interesting.
If you have:
- 1️⃣ A narrow, repeatable task
- 2️⃣ Enough high-quality examples
- 3️⃣ Sufficient volume to justify the training and serving cost
...you can fine-tune a smaller model to perform extremely well on that specific workload, often at a fraction of the inference cost.
🔑 The key is specialization.
You do not need your most capable model handling every request. Use frontier models where broad reasoning matters, and specialized models where the task is predictable and you have the data to train for it.
This is why model selection increasingly looks less like choosing one best model and more like building a portfolio of models for different jobs.
💬 What models do you use the most now open weights or proprietary?
Transcript
Generated from the video's captions and lightly edited.
Viktoria, when should we use proprietary versus open weight models? I'll give you an example. You all probably know Cursor, which is very popular IDE platform. So, originally they built their coding agents on proprietary models. So, OpenAI, as we started all becoming very efficient in coding, right? We started utilizing a lot of these coding agents. So, for their subscription model, the cost became quite cost-prohibitive. They realized they need to do something different. So, what they did, they took open weight model Kimi and then they fine-tuned it specifically for coding tasks.
So, they analyzed their prompt response loops and now they optimized it. In this way, they achieved 86% drop on cost per task compared to proprietary model while maintaining the quality. And now they built really intellectual property that they can utilize. What it really comes down to, do you have task where you also have a training data that you can fine-tune the model? Because that's how you will get to much lower cost, much lower latency, and better results.