A Few Prompts to Train Trillion-Parameter Models: Trajectory Brings Post-Training onto a Platform
Trajectory launches a post-training platform, with $60 million in total funding and a new round led by Sequoia at a $300 million valuation. It simultaneously releases two research previews, Continual grading and GS1, attempting to solve evaluation taste and production feedback.
Trajectory today launched a platform aimed at compressing post-training models into a few agent prompts. Officially, users can use a few prompts to access a complete toolchain for training, evaluating, and deploying trillion-parameter models. For those unfamiliar with this layer, post-training refers to continuing training after the base model to make it better fit specific tasks, preferences, and evaluation criteria.
The video is preserved here:
The launch information also includes funding: Trajectory has raised $60 million in total, with the new round led by Sequoia at a $300 million valuation. That number is not small within the toolchain for training models. In the past, post-training often required coordination among several teams—algorithms, data, evaluation, and deployment—and platformization aims to bring the process into a single entry point.
The two research previews released at the same time are more worth looking at.
Continual grading targets the "taste" problem in evaluation. Human scoring has limits. For a model to learn to judge what counts as good, it needs far more feedback than handwritten annotations. The company says it can scale your taste and let the model learn what "good" looks like from a broader range of feedback.
GS1 is an algorithm that lets models learn directly from behavior in production environments. The key is production feedback. What the model does after launch, what the results are, and whether that can become a training signal are unavoidable parts of continual learning.
The final line of the official tweet is: "You know what good looks like, now go build it. Own your own value function with Trajectory." This line states the product selling point very directly: you define what counts as good, and the platform turns it into model capability.
The entry point becomes simpler, but training itself does not automatically become simpler. Data, evaluation criteria, and compute are still scarce. Trajectory has made post-training into a platform. What to watch next is whether, after a few prompts, users can truly get reproducible results.
发布时间: 2026-11-03 18:37