{"schema":"sanctity-signal/0.2","record_type":"signal","id":"nvidia-nemotron-science-e098","domain":"science","drop_id":"signal-e098","occurred_at":"2026-11-16","how_to_use":"This is a structured record of one real-world action and its competing interpretations, from Sanctity Signal. To build content from it (a reel script, a blog, a thread): open on the hook, ground the viewer with plain and observed, walk the dated trail, present the interpretations as live tension between named readings, and close on the before/after shift and what_to_watch. Hard rules: items under established are fact; items under unknown are never presented as fact; the evidence status (verified or reported) is stated, not dropped; sources are cited.","actor":"NVIDIA","action_type":"release of an open science model package with published weights, a 50,000-question chemistry dataset, post-training recipes and 7 named laboratory and university deployments","observed":"We read the release at nvidianews.nvidia.com, dated November 16, 2026, with a GlobeNewswire dateline of 11:00 Eastern Time, which is 16:00 Universal Time, 68 minutes before this window opened. It was not carded in the evening edition, which swept science and reported it empty; this card is dated by its event and says so plainly. NVIDIA introduced Nemotron for Science, packaging scientific applications of the NVIDIA Nemotron 3.5 Super VL open model with a new open chemistry dataset and post-training recipes. The model and the dataset are \"available now on Hugging Face\". The body text gives no parameter count; the Hugging Face repository slug reads 120B-A12B, which indicates about 120 billion total and 12 billion active parameters. The dataset is LitmusBench v0.2, with \"nearly 50,000 questions\" for training and evaluating chemistry reasoning, covering molecular properties and experimental drug discovery data. The model was optimised to run locally on a 128 gigabyte NVIDIA DGX Spark. Named deployments and figures. Lila Sciences used it to design laboratory protocols, \"matching frontier-level performance without requiring additional fine-tuning\"; after post-training with laboratory feedback, quantum-dot accuracy reached 61% against 54% for GPT-6 Astra, at \"10-20x lower inference cost\". Terray Therapeutics post-trained it to pull reactants, products and conditions out of scientific diagrams, recovering about 76% of annotated reactions on the RxnScribe test and reportedly surpassing published results on 2 benchmarks; no baseline figure is given. Therna post-trained it for mRNA therapeutic design, \"nearly doubling accuracy in predicting protein-expression ranges over frontier models\", with scientists preferring its explanations to the base model's in 96% of comparisons. Argonne National Laboratory, part of the United States Department of Energy, is \"integrating Nemotron into its StructBioReasoner multi-agent system\" for protein-based therapeutic candidates. The Julich Supercomputing Centre is putting it into drug and scientific discovery agents. KTH Royal Institute of Technology built an agent that repairs broken GROMACS molecular dynamics setups and \"fixed 81% of test cases with reasoning enabled, and every fix was checked independently by GROMACS itself\". RIKEN made it available on its RIKYU platform through its RiVault gateway. Not established: the licence, whether \"open\" means freely commercial, and any figure NVIDIA measured itself rather than reporting from a partner.","corrections":[],"evidence":{"status":"verified","note":null,"sources":[{"role":"primary","label":"NVIDIA, Nemotron for Science to accelerate discoveries with agents, Nov 16, 2026","url":"https://nvidianews.nvidia.com/news/nvidia-introduces-nemotron-for-science-to-accelerate-discoveries-with-agents"},{"role":"primary","label":"GlobeNewswire copy of the release with the 11:00 Eastern dateline, Nov 16, 2026","url":"https://www.globenewswire.com/news-release/2026/11/16/3403880/0/en/nvidia-introduces-nemotron-for-science-to-accelerate-discoveries-with-agents.html"},{"role":"corrob","label":"The Next Web on Noeri's $110 million launch with NVIDIA as a strategic investor, Nov 16, 2026, 18:17 UTC","url":"https://thenextweb.com/news/noeri-110m-stanford-yejin-choi-carlos-guestrin-menlo-nvidia"}]},"strength":{"cost":"Medium","commitment":"Medium","reversibility":"Moderate","persistence":"Years","rationale":"Cost is medium. Publishing weights and a 50,000-question dataset gives away work NVIDIA could have sold, but it sells the hardware the weights run on, so the giveaway is a demand generator rather than a sacrifice. Commitment is medium: downloaded weights cannot be recalled, and 4 public research institutions have now built systems on them, which is a harder commitment to reverse than a product launch. Reversibility is moderate, because NVIDIA controls the next version, the licence and the DGX Spark the model was tuned for. Persistence is years: a published model file with a dataset outlives the company's interest in it, and the institutions integrating it will keep running it long after the announcement."},"shift":{"prior":"Frontier science work means paying a laboratory for its closed model.","action":"We read the release at nvidianews.nvidia.com, dated November 16, 2026, with a GlobeNewswire dateline of 11:00 Eastern Time, which is 16:00 Universal Time, 68 minutes before this window opened. It was not carded in the evening edition, which swept science and reported it empty; this card is dated by its event and says so plainly. NVIDIA introduced Nemotron for Science, packaging scientific applications of the NVIDIA Nemotron 3.5 Super VL open model with a new open chemistry dataset and post-training recipes. The model and the dataset are \"available now on Hugging Face\". The body text gives no parameter count; the Hugging Face repository slug reads 120B-A12B, which indicates about 120 billion total and 12 billion active parameters. The dataset is LitmusBench v0.2, with \"nearly 50,000 questions\" for training and evaluating chemistry reasoning, covering molecular properties and experimental drug discovery data. The model was optimised to run locally on a 128 gigabyte NVIDIA DGX Spark. Named deployments and figures. Lila Sciences used it to design laboratory protocols, \"matching frontier-level performance without requiring additional fine-tuning\"; after post-training with laboratory feedback, quantum-dot accuracy reached 61% against 54% for GPT-6 Astra, at \"10-20x lower inference cost\". Terray Therapeutics post-trained it to pull reactants, products and conditions out of scientific diagrams, recovering about 76% of annotated reactions on the RxnScribe test and reportedly surpassing published results on 2 benchmarks; no baseline figure is given. Therna post-trained it for mRNA therapeutic design, \"nearly doubling accuracy in predicting protein-expression ranges over frontier models\", with scientists preferring its explanations to the base model's in 96% of comparisons. Argonne National Laboratory, part of the United States Department of Energy, is \"integrating Nemotron into its StructBioReasoner multi-agent system\" for protein-based therapeutic candidates. The Julich Supercomputing Centre is putting it into drug and scientific discovery agents. KTH Royal Institute of Technology built an agent that repairs broken GROMACS molecular dynamics setups and \"fixed 81% of test cases with reasoning enabled, and every fix was checked independently by GROMACS itself\". RIKEN made it available on its RIKYU platform through its RiVault gateway. Not established: the licence, whether \"open\" means freely commercial, and any figure NVIDIA measured itself rather than reporting from a partner.","update":"A chipmaker gives the weights away and names who is using them."},"inference":{"established":["Weights and the LitmusBench v0.2 dataset are published on Hugging Face now.","Argonne, Julich, KTH and RIKEN are named as integrating the model.","Lila reports 61% against 54% for a frontier model on quantum dots."],"unknown":["What the licence permits. The release names no terms.","That the figures are independent. Each comes from the partner.","That it matches frontier models generally. The tests are narrow."]},"narrative":{"hook":"NVIDIA published a free science model, and a national laboratory has already built on it.","plain":"Nemotron for Science went out on November 16 with weights on Hugging Face and a chemistry dataset of nearly 50,000 questions. Argonne, Julich, KTH and RIKEN are named as users. No licence terms appear on the page.","story":["The pattern on this card is price, not capability. Lila Sciences reports 61% against 54% for a frontier model at 10 to 20 times lower inference cost, and KTH ran its agent overnight on 1 desktop machine. If that holds, the constraint on computational science stops being the model budget."],"trail":[{"d":"SEP 7","t":"A drug designed by software lowers 6 separate measures of biological age in the people who took it. Edition 015, the first science card in this ledger."},{"d":"OCT 8","t":"The White House and 11 AI companies pledge $2.4 billion of compute credits to a United States government science programme. Edition 045. Credits for closed models, not weights."},{"d":"OCT 22","t":"Hugging Face publishes an open model that reads DNA and answers questions in English. Edition 059. The same distribution route this release uses."},{"d":"NOV 16","t":"NVIDIA publishes Nemotron for Science at 16:00 Universal Time, with weights, the LitmusBench v0.2 dataset and 7 named deployments. It also announces separate quantum work with IBM and Fermilab."},{"d":"NOV 16","t":"Later the same day, Noeri launches with $110 million from Menlo Ventures, with NVIDIA, Cisco and Snowflake as strategic investors, to build self-improving systems for companies and for scientific hypothesis work."},{"d":"Reacts","r":true,"t":"OpenAI and Anthropic, whose science access runs through paid programmes and grants; Argonne, Julich, RIKEN and KTH, now dependent on one vendor's weights; Lila, Terray and Therna, whose figures are the evidence; the chemistry journals that will have to referee agent-produced work; every funder asked why it is paying per token."}],"glossary":[{"term":"Open weights","def":"The trained model file is published for download, so anyone can run or modify it. It does not by itself say what the licence permits."},{"term":"Post-training","def":"Further training of a published model on a narrow task or a laboratory's own feedback. The recipes here are the instructions for doing it."},{"term":"GROMACS","def":"The standard open-source package for molecular dynamics simulation. Its setups are notoriously fragile, which is why repairing them is a measurable task."},{"term":"Active parameters","def":"The fraction of a model used for each token. 12 billion active out of 120 billion is why it runs on one desktop machine."}]},"judgment":{"prompt":"Which explanation do you believe?","interpretations":[{"id":"a","label":"Open weights reach the laboratories","text":"4 public institutions integrated it at once. Paid science access is being undercut.","abstain":false},{"id":"b","label":"Hardware demand, dressed as science","text":"The model was tuned for a desktop DGX Spark. Free weights sell boxes.","abstain":false},{"id":"c","label":"Partner figures, not measurements","text":"Every number comes from a company with a reason to publish a good one.","abstain":false},{"id":"d","label":"The usual domain-model release","text":"Vendors ship vertical models with reference customers. This is that, in chemistry.","abstain":false},{"id":"e","label":"Too early to call.","text":"The evidence cannot yet separate these.","abstain":true}],"distribution":{"status":"held back until you have made your call","calls":null}},"story_thread":null,"what_to_watch":[{"linked_interpretation":"a","direction":"strengthens","observable_event":"Argonne, Julich, KTH or RIKEN publishes a paper or system description crediting Nemotron 3.5 Super VL, or a peer-reviewed result reproduces one of the partner figures, by June 30, 2027.","window_end":"2027-06-30","state":"unresolved"},{"linked_interpretation":"a","direction":"weakens","observable_event":"The Hugging Face licence turns out to restrict commercial or derivative use, or NVIDIA withdraws the weights or moves the next version behind an API, by March 31, 2027.","window_end":"2027-03-31","state":"unresolved"},{"linked_interpretation":"b","direction":"strengthens","observable_event":"NVIDIA reports DGX Spark or local science-workstation demand as a named growth driver on an earnings call, by March 31, 2027.","window_end":"2027-03-31","state":"unresolved"}]}