From the NVIDIA edition of October 15, 2026
NVIDIA releases open Nemotron 3.5 Super VL and Bolt models for agents from data center to desktop
NVIDIA · Technology vendor / AI platforms · Global (U.S. vendor; Hugging Face / NIM distribution)
- Publisher
- NVIDIA Blog
- Original publication
- 2026-10-15
- Source retrieved
- 2026-10-16
- Event date
- 2026-10-15
What happened
On Oct 15 NVIDIA released Nemotron 3.5 Super VL, an open 120-billion-parameter multimodal model for specialized always-on agents, plus Nemotron 3.5 Bolt 74B-A9B 24GB for local agentic workloads; Bolt E4B 4GB is due Oct 23. Super VL is positioned for instruction following, tool use, multistep workflows and visual understanding of PDFs, charts and webpages, with claimed class-leading speed (up to 5x). Organizations can post-train on their own data and deploy on their own infrastructure. Named customizers include Scale AI (Skydance, EY), Ramp, AI21/Alice, Mercor and others; Salesforce is said to be customizing Nemotron for AI agents that digitize and automate complex government workflows via Missionforce Operations. Distribution: Hugging Face, OpenRouter, ModelScope and NVIDIA NIM; NeMo for customization; Dynamo claimed to serve up to 8x more users with 85% lower latency.
Why it matters
Direct platform signal for SLED buyers evaluating open multimodal agents that can stay on-premises; Salesforce Missionforce government-workflow mention is a vendor claim needing buyer validation, not a measured SLED deployment.
Evidence and measured results
Primary vendor blog dated Oct 15, 2026 by Kari Briski; partner results are vendor-attributed; no independent measurement in this page.
Limitations and uncertainty
Vendor claims for 5x speed, Dynamo 8x/85% and partner outcomes are not independently verified here; government workflow mention is Salesforce customization intent, not a public-sector case study with metrics.
Put this evidence to work
Lighthouse Advisory interpretation, grounded in this source. Enriched 2026-10-16; this does not change the original publication date. Each role follows the same checklist so takeaways compare across sources.
Sales
Role takeaway
- Customer problem
- Agencies need multimodal agents that read documents and tools under their control without sending sensitive content to closed frontier APIs.
- Stakeholders
- CIO / Chief AI Officer
- Mission/application owners
- Security/privacy
- Procurement
- Discovery questions
- Which document and workflow classes must stay on-premises?
- Is Salesforce Missionforce or another integrator already in scope?
- Will evaluation use open weights only or require NIM support?
- Value hypothesis
- Interpretation: open multimodal agents may cut closed-API spend and improve document automation if post-training works on agency data; this source reports no SLED measured results.
- Potential engagement
- Bounded open-weight evaluation on a sample of agency PDFs/charts plus a written production-path decision (open serving vs NIM).
- Unsupported claims
- Guaranteed 5x throughput in agency environments
- That Missionforce government workflows are live with measured outcomes
- That Dynamo 8x/85% applies without buyer testing
Pre-sales engineering
Role takeaway
- Fit
- High for document-agent and local-agent PoCs on NVIDIA GPUs; uncertain for non-NVIDIA inference estates.
- Architecture and integration
- Open weights via HF/vLLM-class stacks or NIM microservice; NeMo post-training; optional Dynamo for multi-GPU serving; Bolt for 24GB local desktops/Sparks.
- Prerequisites
- GPU capacity and drivers
- Document corpus and eval harness
- Identity/network isolation for on-prem inference
- Constraints
- Vendor speed and partner metrics unverified here
- Bolt E4B not available until Oct 23
- Government workflow claim is integrator intent
- Security
- On-prem framing keeps data local; still need RBAC, logging, model-update and prompt-injection controls for agents.
- Proof of value
- Score OCR/table/chart tasks on agency samples
- Compare open-weight latency vs any NIM trial
- Document data residency decision before production
Delivery
Role takeaway
- Work
- Stand up evaluation serving, run document-agent tests, optionally post-train with NeMo, then decide production packaging.
- Dependencies
- GPU ops
- Data owners for training samples
- Security review
- Optional Salesforce/integrator if Missionforce path
- Ownership
- Agency AI platform owner with security co-sign on production.
- Skills and adoption
- Need engineers for NeMo/SFT and inference ops; end users need agent workflow training.
- Governance checkpoints
- Eval charter approved
- Post-train data inventory cleared
- Production license/support path approved
- Acceptance criteria
- Proposed: baseline accuracy on named document tasks before/after any post-train
- Proposed: written decision on open serving vs NIM
- Risks
- Vendor benchmarks overstate end-to-end agent time
- Integrator path bypasses agency architecture review
- License cost surprise at production
Implementation considerations
Lighthouse Advisory interpretation across the operating dimensions a public-sector buyer must settle before this evidence becomes a design. Each note answers the question under its heading for this specific source.
Architecture and integration
What must connect, and where does the AI sit in the workflow?
Open weights plus optional NIM packaging; runs across DGX Spark/Station, RTX PRO/RTX, Jetson, data center and cloud; Bolt fits 24GB local memory with native 2-bit quantization; NeMo for post-training.
Governance
Who approves, reviews and stays accountable for outcomes?
Vendor emphasizes customer control of post-training data, tools and policies; production governance still buyer-owned.
Security and privacy
What data, permissions and controls need testing?
On-premises deployment framed as keeping data private; NIM said to refresh security patches every two weeks—buyers must still assess residency, access control and model-change governance.
Accessibility and workforce
Who is affected, and what skills or accommodations follow?
Local Bolt models may help agencies keep agent work off cloud for staff without large clusters; skills needed for NeMo post-training and NIM ops.
Procurement
What should contracts, pricing and exit terms secure?
Open weights enable evaluation; production NIM/enterprise support path implies separate software licensing (see pricing resource).
Operating model
Which teams own the service once it runs?
Shift from closed API agents to adaptable open models that agencies can post-train and host; requires MLOps ownership.
Publication history
- 2026-10-15NVIDIA · Issue 407 resources
Stable resource ID: nvidia-nemotron-35-super-vl-blog-20261015