Should You Train Your Own Model: The Pros and Cons of Post-Training for Application Companies

As AI application companies pursue a 'full-stack' strategy that integrates into the model layer, whether to train models directly has emerged as an important topic. Rather than Pre-training models from scratch, most companies are focusing on Post-training and RL (Reinforcement Learning) based on powerful Open-weights models.

The training spectrum can be divided as follows:

• Prompt engineering and RAG (no training)
• Fine-tuning small models
• SFT (Supervised Fine-Tuning) and RL based on open-weight models
• Continued pre-training
• Pre-training from scratch

The main reasons to adopt post-training are as follows:

1. Improving Unit Economics and Latency As scale grows, API call costs become a burden. Small models specialized for specific tasks can respond much more cheaply and quickly than frontier models. For example, Intercom's Fin Apex 1.0 lowered costs to about 1/5 the level of frontier models while also cutting response time by 0.6 seconds.

2. Breaking Away from Dependency and Securing Control Relying solely on a specific frontier API exposes companies to risks such as price fluctuations, rate limits, and the API provider itself emerging as a direct competitor. In Cursor's case, despite high usage, it faces risks of recording low gross margins or becoming vulnerable to policy changes by API providers.