Turning Proprietary Data Into Sovereign Intelligence: Post-Training Open-Source Models Achieve 40% Higher Performance Than GPT 5.4 Nano

At an NVIDIA event, Jensen Huang said: With customization and ecosystem partners, these models can become frontier models, and proprietary data can become sovereign intelligence.

Baseten reposted this and thanked the NVIDIA team for mentioning its work. The specific thing mentioned was this: Baseten and the CodeRabbit team worked together to post-train an open-source model and bring it to frontier-level capability.

The slide in the video gives two key pieces of information. On the left are the CodeRabbit logo and the code editor interface; on the right is a comparison of two bars: GPT 5.4 Nano and Nemotron 3.5 Lightning. The latter has 40% higher performance. The subtitle at the bottom reads “Here's Baseten helping CodeRabbit.”

For readers unfamiliar with this field, a quick bit of background: post-training is not training a model from scratch; it means continuing to feed data and tune parameters on an open-source base so that it performs better on a specific task. CodeRabbit does code review, a scenario that involves large amounts of private code and team conventions, which general-purpose models find hard to fully cover. Using your own data for post-training is equivalent to giving a general-purpose model a push in a specialized direction.

The term “sovereign intelligence” sounds grand, but in engineering terms it is actually not complicated: model capabilities are built on data and infrastructure that you control yourself, rather than relying on the black box of an external API. For teams that treat their codebase as a core asset, this matters more than benchmark scores.

The 40% figure comes from NVIDIA's presentation materials; the specific test set and evaluation method were not elaborated. But it at least shows one thing: open-source models plus post-training can already be compared with closed-source general models in vertical scenarios. Not every team needs to train a model from scratch; taking an open-source base and post-training it with your own data may be a more realistic path.