@nvidia’s Nemotron 3 Ultra handles software-engineering tasks at a fraction of the per-task cost of frontier models. So we trained a router to send each coding task to the cheapest model that can successfully solve it, cutting inference cost while holding frontier-level quality. The result: GPT-5.5-level pass rates on held-out SWE-bench Verified at ~25% lower cost. The oracle policy sends 72% of tasks to Nemotron 3 Ultra, and the trained router captures most of that.

The models are complementary. The trained router sends 73% of tasks to @NVIDIAAI's efficient Nemotron 3 Ultra and routes the long tail to GPT 5.5 and Opus 4.7 on tasks where frontier performance at a premium is worth the tradeoff. Since the router is agentic, it can call tools in a sandbox with the task codebase to match the task against each of the candidate models’ strengths and weaknesses.

Furthermore, we find that the token-level granularity in self-distillation is both a strength and a weakness – it's dense, but there is often a great deal of noise in the updates because the teacher and student may disagree on tokens for reasons unrelated to the desired behavior. We want to concentrate our loss on the tokens which matter most for the student's improvement. This is the intuition behind RMSD: we filter for a set of T token positions where the teacher and student disagree the most, then an LLM judge narrows those to a final set of S tokens that are most relevant to the target behavior. In our experiments, RMSD reaches the target behavior in about half the steps of OPSD, making it significantly more data efficient, while also training more stably and reaching a higher performance ceiling. We think that the self-distillation approach and RMSD in particular will be powerful methods for advancing the capabilities of models on specialized settings.

Some enterprise tasks are challenging to hill-climb with RL-based methods since they involve very out-of-distribution behavior. On-policy self-distillation (OPSD) gives a model learning signal for every token it writes, far richer than the single scalar reward of RL. But that channel is noisy: most tokens don't reflect the behavior you're after. We introduce Relevance-Masked Self-Distillation (RMSD), which uses a two-step filtered loss mask to cut through the noise and find the tokens with the highest signal. Compared to OPSD it trains more stably, provides higher data efficiency, and reaches a higher performance ceiling.