Coinbase has also trained its own small model internally, based on quwen3.5. So, more vertical business companies are starting to train their own small models in-house, possibly to save costs or for privacy data considerations. This demand is likely to grow, and inference platforms in the market like Fireworks, together.ai, and Baseten have also acquired tech teams to expand their post-training platform businesses. There are also post-training platforms like Thinking Machines Lab and Bespoke Labs working in this post-training market. Additionally, another type of company is emerging in this space, such as Snorkel AI and Bespoke, which are more focused on data and RL environment post-training infrastructure. So, the future clients of these companies likely won’t be limited to top-tier labs but will also include vertical model companies purchasing their environments and data. In fact, companies like River AI, Mind Lab, and Predibase, which operate under the banner of continuous learning and efficiently serve a large number of specialized LoRA models on shared GPUs, are essentially also part of this post-training market.