Nemotron targets locally controlled science
NVIDIA has used its SC26 address to introduce Nemotron for Science and demonstrate AI agents working directly with scientific software. The toolkit combines the Nemotron 3.5 Super VL open model with an open chemistry dataset and post-training recipes for life sciences and chemistry. Researchers can run it on their own workstation, DGX Spark or supercomputer. The local-deployment emphasis is important for laboratories handling unpublished data, controlled methods or compute environments that cannot send every artefact to a hosted service.
At Argonne National Laboratory, NVIDIA says an agent received a single prompt, planned and tested hypotheses, selected simulations and identified two new binding sites. That workflow goes beyond summarising papers or generating code snippets. The agent participates in an iterative scientific loop, choosing which expensive experiment to run next based on intermediate results. Such autonomy can expand the number of hypotheses explored, but it also makes experimental policy, stopping conditions and provenance part of the software design.
NVIDIA highlighted an agent from KTH Royal Institute of Technology that repairs broken GROMACS molecular-dynamics setups. In an overnight run on one DGX Spark, it reportedly fixed 81 per cent of test cases. Repair agents could save specialists from repetitive configuration work and make simulation tools more accessible. The unresolved 19 per cent is just as important: teams need to know whether failures are detected, silently produce invalid physics or require an expert to reconstruct what the agent changed.
Agents move from suggestions to experiments
The new Nsight AI Porting Agent provides a more concrete software-engineering example. NVIDIA says it ported a routine from the Weather Research and Forecasting model to GPUs and achieved a speed-up of more than five times. The task consumed about 150 million tokens. That result demonstrates persistence and tool use on a complex codebase, while also showing that agentic porting can be compute-intensive. Economics should be measured against expert engineering time, energy and the useful lifetime of the optimised routine.
A speed-up is only valid if numerical behaviour remains acceptable. Weather and scientific models can be sensitive to floating-point ordering, parallel reductions and approximation choices. Ported code needs tests against trusted outputs across representative inputs, not just a successful compilation or one benchmark. Researchers should specify tolerances, conservation properties and reproducibility expectations before the agent begins. Where exact equivalence is impossible, the change and its scientific implications must be documented for downstream users.
Porting results show promise and high token cost
NVIDIA also described scientific agents operating telescopes and simulation environments, reinforcing the role of CPUs between model calls. An agent loop may parse a result, invoke a compiler, prepare a new simulation and evaluate output before the GPU performs the next heavy computation. Infrastructure planning must therefore balance accelerator, CPU, memory, storage and network capacity. An idle accelerator waiting on orchestration is no more efficient than a human workflow blocked by slow simulation.
Running the model locally improves data control but does not automatically secure the workflow. Agents may ingest untrusted repositories, papers or data files containing instructions designed to alter tool use. Scientific environments need isolated execution, approved dependencies and restrictions on network and file-system access. Instruments and production clusters should not be reachable with the same authority used in a disposable test. Every generated patch, configuration and experiment should retain a signed lineage to the model and source inputs.
Reproducibility must scale with automation
Open datasets and post-training recipes make independent evaluation more feasible. Research groups can test the toolkit on their own domains and inspect whether gains transfer beyond NVIDIA’s demonstrations. They should publish unsuccessful cases as well as successes, particularly where the agent creates plausible but invalid explanations. Benchmarking should include total token and energy use, expert review time, reproducibility and the number of scientifically useful results—not merely how many tasks the agent attempted overnight.
Nemotron for Science represents a meaningful shift from general assistants toward specialised agents that operate research tools. The strongest promise is not autonomous discovery without scientists, but automation of the iterative work between a hypothesis and a validated result. NVIDIA’s examples show real potential and real cost. Institutions should begin with bounded, replayable workflows where errors are detectable, then expand autonomy only when the agent’s changes, calculations and decisions remain understandable to the experts accountable for the science.
Scientific institutions will also need a new review discipline for agent-authored code. Conventional peer review examines methods and results, while agentic workflows add prompts, tool versions, generated patches and branching decisions that may affect reproducibility. Those artefacts should travel with a paper or internal result where they materially influenced the computation. Long-running agents can explore many failed approaches before succeeding, and retaining every token may be impractical. Teams should define which checkpoints and rationales are essential, hash large intermediate datasets and preserve the final executable environment so another researcher can verify the claimed improvement without repeating the entire costly search.