NVIDIA drops 6M multilingual reasoning examples, cutting edge model costs 60%

NVIDIA isn’t just releasing models anymore—it’s open-sourcing the raw intelligence behind them. The company just published the Nemotron Post-Training Dataset V2, a collection of 6 million reasoning examples translated into French, Spanish, German, Italian, and Japanese. This follows their v1 release, which powered the well-received Llama Nemotron Super model, and it represents a deliberate strategy to arm the open ecosystem with not just weights, but verifiable training data.

The translation pipeline itself is worth noting. NVIDIA didn’t just throw an LLM at the problem. Recognizing that open-source models “deteriorate significantly as input length increases” and hallucinate more on structured SFT data than on standard test sets like FLORES, the team built a clever, paranoid pipeline. They broke text down by newlines, wrapped translations in special brackets for clean extraction, and used fastText language ID to filter out off-target results—discarding over 55,000 examples in the process. For the actual work, they benchmarked and selected Qwen2.5-32B-Instruct for German and Qwen2.5-14B-Instruct for the other languages, specifically because the models could fit on a single A100 GPU and carried Apache 2.0 licenses.

This dataset release coincides with the launch of the Nemotron Nano 2 9B, a model built for the edge that uses a hybrid Transformer–Mamba architecture. The real story there is the “thinking budget”—a configurable knob that lets developers control how many reasoning tokens the model uses. NVIDIA claims this can slash reasoning costs by up to 60% while delivering up to 6× higher token generation throughput compared to peers in its size class. That’s a direct shot across the bow of models that burn compute on unnecessarily long chains of thought.

The practical audience here is crystal clear: developers building customer service agents, support chatbots, and analytics copilots who need a model that runs efficiently on edge devices or RTX hardware. By preserving the original English reasoning chains and only translating the user prompt and model response, NVIDIA is betting that pre-training knowledge in English is the best foundation, and language is just a surface interface. It’s a pragmatic admission that while multilingual models are improving, the deepest reasoning still happens in English. The model weights are on Hugging Face, and it’ll be available as a NVIDIA NIM microservice for those who want high-throughput, low-latency inference without managing infrastructure.