Baseten joins NVIDIA's Nemotron Coalition
Baseten has joined NVIDIA's Nemotron Coalition and will contribute post-training environments and inference expertise to NVIDIA's open models, as more enterprises customize open weights for production use.
Inference provider Baseten announced that it has joined the NVIDIA Nemotron Coalition. The move expands its existing work with NVIDIA on the Nemotron family of open models. Baseten says it will feed lessons from enterprise workloads, along with its post-training and inference expertise, back into Nemotron development.
The two companies already work together. Baseten was a day-zero launch partner for Nemotron Super 3.5 VL and Nemotron 3 Ultra. It also built reinforcement learning environments aimed at instruction following, which it described as a weak spot for Nemotron. Other environments cover long-horizon tasks, including statistical analysis of biological datasets, theorem proving and competitive game playing. Baseten also trained vision adapters for Nemotron 3 Ultra, GLM-5.2 and GLM-5.3, and reported "significant improvements" on the MMMU-Pro benchmark. It did not give figures.
The company pointed to two customers:
- CodeRabbit fine-tuned Nemotron 3.5 Lightning for code-review routing using NVIDIA NeMo tools on Baseten Training. On a fixed set of 1,000 tasks, agreement with its routing policy rose from 75.8% to 80.7%, and inference cost at saturated throughput fell 50.4%.
- Harvey post-trains Nemotron on Baseten for legal workflows that it says serve tens of thousands of lawyers daily.
Baseten also said a single NVIDIA Hopper GPU running Nemotron 3.5 ASR Streaming handled up to 100 concurrent real-time audio streams in its own benchmark. Its Baseten Loops product, which links managed RL to dedicated inference, now supports Nemotron 3.5 Super VL and Nemotron 3 Ultra. Customers keep ownership of the weights they train.
The post does not say what coalition membership involves, what it costs, or which other companies belong.