The last time I watched a technical interview with @ FireworksAI-HQ and Curor was at https://(x.com)/sonyatweetybird/status/2059323965219491937, When it comes to Curor's model strategy and why one makes their own model decisions: 1) The cost of a generic model is higher than that of a proprietary model, and the generic model also uses domain data for proprietary training (such as coding domain) during training. Curor accumulates a large amount of user behavior data. 2) The strategy of Curor+FireworksAI platform is top-down (starting with user applications, training based on open source models, and then building independent pre trained complete models), while the model company is bottom-up (from model development to scenario/environment applications, from open model capabilities to data feedback loops). Curor should take the final step and collaborate with Xai's computing power to create its own pre training model. The latest customer success story of @ FireworksAI-HQ, on the other hand, is related to the largest legal application @ harvey, which trains its own legal proprietary model based on an open-source model. The core is that the cost is lower than that of a general model, but the ability is close to that of a general model. They found that: 1) Mixed legal representation can surpass cutting-edge models in terms of quality and cost by selectively routing tasks to cutting-edge advisors. We tested a hybrid setup where GLM 5.1 serves as the primary working node and routes tasks to Opus 4.7 as a consultant when needed. GLM calls Opus very few times, with an average of only 0.83 calls per task. The quality and cost of hybrid settings are superior to Opus: in the same 100 tasks, the average pass rate of hybrid settings is 18%, while Opus is 14%; The costs are $368 and $954 respectively. 2) Post training processing can improve the performance of open-source models to the level of cutting-edge legal models. In the 100 task slices of the Legal Agency Benchmark Test (LAB), SFT increased the average pass rate of Kimi 2.6 from 11% to 15%, surpassing Opus' 14%. But the cost gap is even more significant: in the same 100 tasks, the cost of the hybrid setup is only $84, while Opus is $954, reducing the cost by about 11 times. While the general model is constantly cannibalizing the application scenarios, it is applied vertically to balance the cannibalization of the general model on the basis of non single models (routing of costs and capabilities)+vertical domain data+top-down (such as curor's strategy).