Baseten Brings Inference Expertise to NVIDIA’s Nemotron Coalition

Baseten announced on November 5, 2026 that it has joined the NVIDIA Nemotron Coalition, contributing insights from enterprise workloads and expertise in post-training and inference to open-model development. The membership expands its NVIDIA collaboration across model launches, reinforcement learning, customization and dedicated inference.

NVIDIA launched the coalition on March 16, 2026 to pool research, expertise, data and compute around open frontier models. NVIDIA’s launch announcement said the first project would be a base model co-developed by Mistral AI and NVIDIA, trained on NVIDIA DGX Cloud, released as open source for specialization by developers and organizations, and used as the foundation for the upcoming Nemotron 4 family.

Baseten was a day-zero launch partner for Nemotron 3.5 Super VL and Nemotron 3 Ultra. Its coalition role extends beyond serving finished models to training environments and operational knowledge shaped by enterprise use. The practical focus is customization for specific business requirements, with deployment quality, cost and flexibility treated as connected concerns.

Training for Production Workloads

The company developed reinforcement-learning environments for complex instruction hierarchies, ignoring extraneous information and following detailed style preferences. Other environments address long-horizon reasoning and planning through statistical analysis of complex biological datasets, mathematical theorem proving and competitive game playing. Those tasks address the way production instructions arrive through system prompts, tool definitions, user follow-ups, tool responses and error messages.

Its multimodal work adds visual and spatial reasoning tasks and multi-turn tool calls that require visual understanding. Baseten reports that post-training improved results on MMMU-Pro, an expert-level multimodal reasoning benchmark. A related paper submitted on October 5, 2026 describes a 50-million-parameter projector connecting Kimi K2.6’s vision encoder to GLM 5.2 and GLM 5.3 while keeping pretrained weights fixed, with evaluations on MMMU-Pro and BLINK.

Measured Customization Results

CodeRabbit’s August 11, 2026 research account describes a project with NVIDIA and Baseten that customized Nemotron 3.5 Lightning for code-review routing through supervised fine-tuning followed by reinforcement learning with verifiable rewards. On a frozen 1,000-task evaluation, exact route agreement moved from 75.8% for the baseline to 80.4% after supervised fine-tuning and 80.7% after reward training. Cohen’s kappa, a measure of agreement beyond chance, rose from 0.429 to 0.461 and then 0.544.

CodeRabbit treated the final 0.3-percentage-point route-agreement change as non-regression, rather than a statistically decisive improvement, because its paired confidence interval crossed zero. Its serving comparison estimated a 50.4% cost reduction at saturated throughput: the same evaluation cost $2.34 on the baseline model and $1.16 on one Baseten A100. Sustained-load and production-traffic checks remained for a later phase.

Harvey offers a different customization case. Baseten reports that the legal AI platform post-trains Nemotron on Baseten for specialized workflows serving tens of thousands of lawyers daily. Separate research published by Harvey and Baseten on September 8, 2026 loaded an M&A data room into a Python environment and let a root agent delegate bounded review to sub-agents. Harvey reported that across seven models on 50 held-out data rooms, the harness raised mean rubric-criteria pass rates from 23.3% to 62.4%.

Harvey’s account also reports that reinforcement learning on a Qwen3.5-122B-A10B root model increased its pass rate from 29.9% to 63.0% on those 50 rooms, while describing the experiments as initial explorations and the remaining gap to near-perfect pass rates as substantial.

Deployment and Voice AI

Baseten Loops connects managed reinforcement-learning workflows to Dedicated Inference, while customers retain ownership of their trained model weights. Loops supports Nemotron 3.5 Super VL and Nemotron 3 Ultra, with Super VL available for BF16 and NVFP4 training.

The companies are also evaluating Nemotron-powered voice agents across complete applications, including model quality and responsiveness. Baseten’s August 4, 2026 speech-recognition benchmark reported that one NVIDIA H100 sustained up to 100 concurrent real-time WebSocket audio streams. Finalization latency, the time from a speaker finishing to the final transcript being ready, rose from 98 milliseconds at eight streams to 138 milliseconds at 100.

Baseten’s coalition membership connects model adaptation with the infrastructure used to serve the resulting models. The customer results provide concrete evidence for that approach, while CodeRabbit’s saturated-throughput estimate and unfinished production checks, together with Harvey’s early-stage experiments, define its limits. The central question is how well the measured gains hold under sustained production use.