{"resourceId":"nvidia-nemotron-35-super-vl-devblog-20261015","versions":[{"version":"external-4677bed6a1f33af14140bfab0e0bd48b614df60fd86cd4d7b174f5b185f21b93","resource":{"id":"nvidia-nemotron-35-super-vl-devblog-20261015","title":"Developer blog: Nemotron 3.5 Super VL MoE activates 12B parameters; NeMo post-training and NIM for production","organization":"NVIDIA","sector":"Technology vendor / AI platforms","geography":"Global","publishedAt":"2026-10-15","publicationDate":"2026-10-15","eventDate":"2026-10-15","sourceName":"NVIDIA Technical Blog","sourceLabel":"Vendor technical blog","sourceUrl":"https://developer.nvidia.com/blog/nvidia-nemotron-3-5-super-vl-powers-specialized-agents-with-multimodal-reasoning/","evidenceClass":"vendor-claim","outcomeClass":"emerging","topics":["developers-agents","infrastructure","knowledge-work"],"finding":"The Oct 15 technical blog details Nemotron 3.5 Super VL as a 120B-parameter MoE that activates 12B parameters per token, combining reasoning, instruction following, tool use and text/image/document/video understanding. Vendor-attributed partner results include Scale AI for Skydance vendor-credit processing (claimed 40x lower inference cost and 5x lower latency), Ramp receipt-to-JSON (up to 50% higher throughput and 10x lower latency), and AI21/Alice cybersecurity post-training (41% accuracy improvement). Hermes Bench 63.44% and LangChain evals 87.05% are reported versus Nemotron 3 Ultra. Enterprise support for NIM is through NVIDIA AI Enterprise. Training notes include Indic data with IIT Madras/AI4Bharat and open datasets on Hugging Face.","sledRelevance":"Useful engineering detail for SLED platform teams sizing MoE inference and post-training; partner metrics remain vendor-attributed commercial examples, not public-sector evaluations.","evidence":"Primary NVIDIA technical blog Oct 15, 2026 by Chris Alexiuk and Chintan Patel; includes tables and partner figures without independent protocols.","architectureImplications":"MoE 12B-active inference; NeMo AutoModel SFT/LoRA and NeMo RL/Gym; NIM packaging; Dynamo KV-aware routing; NVFP4 quantization and LiLiCorr speculative decoding on DGX Spark.","governanceImplications":"Post-training on org data/tools/policies is the stated path to specialized agents; buyers must govern training data rights and model change control.","securityPrivacyImplications":"Cybersecurity partner example is attack-resilience post-training; still not a substitute for agency red-team and data-handling policy.","caveats":"Partner cost/latency/accuracy figures lack published task definitions and sample sizes in this page; AI-generated summary banner notes possible incompleteness.","streamIds":["nvidia"],"roles":{"sales":"Sales framing: the developer blog supplies MoE sizing (120B total, 12B active) and vendor partner metrics that commercial buyers will cite—40x cost, 10x latency, 41% accuracy—but those are specialized post-trains, not base-model guarantees. Interpretation: use the architecture facts for sizing conversations and treat every percentage as a hypothesis to reproduce. For SLED, emphasize that enterprise support for NIM is via AI Enterprise, so software cost must enter the commercial discussion early.","engineering":"Engineering framing: Super VL is a multimodal MoE activating 12B parameters per token with NeMo SFT/LoRA/RL paths, NVFP4, LiLiCorr and Dynamo recipes for multi-GPU serving. Pre-sales engineers should design PoCs around agency document and tool traces, set GPU memory and quantization assumptions explicitly, and keep NIM versus open vLLM-class serving as alternative deployment modes. Do not accept partner 40x/10x/41% figures without matching task definitions and baselines.","delivery":"Delivery framing: implementing Super VL means standing up NeMo post-training, an evaluation harness for Hermes-like agent behaviors and document tasks, then choosing NIM with AI Enterprise support or open serving. Delivery should time-box a post-train experiment on a narrow workflow, publish before/after metrics with sample size, and gate production on security review of tool use. Acceptance criteria below are proposals, not observed results."},"roleDetails":{"sales":{"customerProblem":"Buyers need credible sizing and customization paths for multimodal agents without treating vendor partner charts as guaranteed savings.","stakeholders":["Enterprise architects","AI platform leads","Finance for license TCO","Security"],"discoveryQuestions":["Which partner-style workflow is closest to ours?","What baseline cost/latency do we measure today?","Is AI Enterprise already entitled on our GPUs?"],"valueHypothesis":"Interpretation: MoE efficiency plus post-training may improve specialized agents; this source's partner percentages are vendor-attributed.","potentialEngagement":"Architecture workshop plus a single post-train PoC scoped to one document or security workflow with buyer-owned metrics.","unsupportedClaims":["That base Super VL beats Claude/GPT generally","Guaranteed 40x inference cost reduction","That Hermes/LangChain scores transfer to agency tasks"]},"engineering":{"fit":"Strong for NVIDIA-centric agent platforms needing multimodal MoE; weaker where CUDA stack is unavailable.","architectureIntegration":"NeMo post-train then open or NIM serve with optional Dynamo multi-node; NVFP4 on Blackwell with fallback paths noted in ecosystem recipes.","prerequisites":["NeMo tooling","Labeled or preference data","Eval harness","GPU operators"],"constraints":["Partner metrics lack protocols here","Indic/training details are research context, not SLED requirements","NIM support needs AI Enterprise"],"security":"Tool-using agents need sandboxing and audit; cybersecurity post-train example is not a control baseline.","proofOfValue":["Reproduce one partner-like task on agency data","Report active-parameter memory footprint","Compare NIM vs open serving support model"]},"delivery":{"work":"Build post-train pipeline, run controlled evals, document deployment mode and support entitlement.","dependencies":["Data labeling capacity","GPU cluster","NVIDIA support channel if NIM","App owners for tool APIs"],"ownership":"AI platform engineering with application owner for workflow acceptance.","skillsAdoption":"ML engineers for NeMo RL/SFT; platform SRE for NIM/Dynamo.","governanceCheckpoints":["Training-data approval","Eval protocol signed","Production support entitlement confirmed"],"acceptanceCriteria":["Proposed: documented before/after on buyer task with n at or above agreed sample","Proposed: chosen serving mode with rollback"],"risks":["Overfitting to narrow demos","Support gap if open serving assumed free forever","Eval leakage from vendor charts"]}},"retrievedAt":"2026-10-16T02:56:30Z","enrichedAt":"2026-10-16T02:57:15Z","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Open cookbooks and datasets lower barrier for research labs; production still needs MLOps skills.","procurementImplications":"NIM enterprise support explicitly routed through NVIDIA AI Enterprise licensing.","operatingModelImplications":"Specialized agents require ongoing post-training and evaluation loops, not one-time model download.","sourceVerification":{"openedUrl":"https://developer.nvidia.com/blog/nvidia-nemotron-3-5-super-vl-powers-specialized-agents-with-multimodal-reasoning/","referenceExcerpt":"architecture activates 12B parameters per token, balancing model capacity with inference efficiency.","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}