From the NVIDIA edition of October 23, 2026
NVIDIA releases Nemotron 3.5 Bolt E4B 4GB, an open small model post-trained from Gemma 4 E4B for local conversational agents on RTX, Jetson, RTX Spark and DGX Spark
NVIDIA · Technology vendor (open models for local and edge AI) · Global
- Publisher
- NVIDIA Research (Nemotron)
- Original publication
- 2026-10-23
- Source retrieved
- 2026-10-24
- Event date
- 2026-10-23
What happened
NVIDIA released Nemotron 3.5 Bolt E4B 4GB, an open small language model built on Google's Gemma 4 E4B with Nemotron post-training, aimed at voice and device assistants, interactive characters and robotic or IoT interfaces running on GeForce RTX, Jetson, RTX Spark and DGX Spark. NVIDIA claims about 40% higher average quality than its prior model on conversational-agent benchmarks, up to about 80% fewer output tokens than peers, operation in as little as about 4 GB in target workloads, multi-turn tool use, configurable reasoning, 64k context and support for 35 languages. It cites memory, latency, cost and privacy constraints as reasons developers choose local models.
Why it matters
Limited and indirect: small local models could support on-device assistants for public-facing kiosks, accessibility tools or field devices where data cannot leave the device, and teaching on campus RTX or Jetson hardware. No public-sector deployment is described.
Evidence and measured results
Vendor release with vendor-selected benchmarks measured on an RTX 5090; no independent evaluation on the page. See the Atomic Agent test for a small independent comparison.
Limitations and uncertainty
Vendor claims; benchmarks chosen by NVIDIA; no SLED deployment evidence.
Put this evidence to work
Lighthouse Advisory interpretation, grounded in this source. Enriched 2026-10-24; 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 want assistants that run on local devices without sending data to the cloud.
- Stakeholders
- Agency CIO
- Accessibility coordinator
- Privacy officer
- Field-operations lead
- Discovery questions
- Which tasks must run without network access?
- What device hardware is available?
- What licence terms are acceptable?
- How will tool actions be authorised?
- Value hypothesis
- Interpretation: local small models could reduce data exposure for simple assistant tasks; unproven in public sector.
- Potential engagement
- Local-model evaluation on agency tasks and hardware.
- Unsupported claims
- NVIDIA benchmark gains apply to agency workloads
- Fits in 4 GB for all uses
Pre-sales engineering
Role takeaway
- Fit
- Edge or desktop devices with NVIDIA GPUs needing a small conversational agent.
- Architecture and integration
- Local inference via llama.cpp, Ollama or similar on RTX/Jetson/DGX Spark; tool calls to local APIs.
- Prerequisites
- Compatible GPU
- Agent framework support for the chat template
- Licence review
- Constraints
- 64k context limit
- Vendor-only benchmarks
- Small-model accuracy limits
- Security
- Restrict tool permissions and log actions; local execution limits data egress but not misuse of tools.
- Proof of value
- Compare with Gemma 4 E4B on agency tasks
- Measure memory and latency on target devices
Delivery
Role takeaway
- Work
- Evaluate, integrate and pilot the model with guardrails.
- Dependencies
- Framework compatibility
- Licence approval
- Test datasets
- Ownership
- Agency innovation or IT team.
- Skills and adoption
- Staff need skills in local model serving and evaluation.
- Governance checkpoints
- Licence review
- Privacy assessment
- Pilot evaluation
- Acceptance criteria
- Proposed: meets task accuracy threshold on agency test set
- Proposed: runs within device memory limits
- Risks
- Overstated vendor claims
- Template or framework incompatibility
- Unsafe tool actions
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?
Supports local inference on consumer and edge GPUs, keeping data on device; 64k context limit.
Governance
Who approves, reviews and stays accountable for outcomes?
Derived from another vendor's base model; licence terms and model provenance need review before public-sector use.
Security and privacy
What data, permissions and controls need testing?
Local execution can keep prompts and data on device, but tool-calling agents still need permission controls.
Accessibility and workforce
Who is affected, and what skills or accommodations follow?
Speech-ready conversational mode and multilingual support could matter for accessible public services; untested in that setting.
Procurement
What should contracts, pricing and exit terms secure?
Open weights reduce licence cost but add review of licence and support arrangements.
Operating model
Which teams own the service once it runs?
Teams adopting the model would own evaluation, guardrails and updates locally rather than relying on a hosted service.
Publication history
- 2026-10-23NVIDIA · Issue 486 resources
Stable resource ID: nvidia-nemotron-35-bolt-e4b-4gb-local-agent-model-20261023