A model’s product name is a poor substitute for a deployment budget. Save the exact file and runtime settings before deciding whether it fits your device.
The smaller Bolt model is now released
NVIDIA’s October 23 release introduces Nemotron 3.5 Bolt E4B 4GB for local conversational agents. The model card lists 8B total parameters, 4B effective parameters, and a 64K context limit.
The quantized checkpoint uses Q4_K_M and is governed by OpenMDW 1.1. A separate BF16 checkpoint and optional DFlash drafter are different artifacts with different deployment requirements.
NVIDIA reports improvements on its selected conversational and agent benchmarks. We have not independently reproduced those quality or speed comparisons.
The download and the runtime measure different things
The repository file metadata lists the GGUF at 6,180,446,208 bytes, approximately 6.18 GB. This is its stored file size.
The project’s runtime measurements depend on hardware, context, cache, and inference settings. Its published examples use a particular quantization and batch size.
A stored file size doesn’t indicate how much memory a GPU needs or how much the whole process allocates. Keep those measurements separate when reviewing the 4GB product name.
Budget the complete local workflow
Record a short deployment manifest before testing a voice or device assistant. Include every model that must remain loaded alongside Bolt.
- Record the checkpoint filename and revision.
- Separate CPU memory from GPU memory in the observed runtime.
- Include the actual context and cache settings.
- Account for a drafter, speech model, or other adjacent component.
- Measure startup and peak memory during a representative conversation.
Then test a task that needs both dialogue and a tool action. A fluent response doesn’t prove the correct tool ran or that its arguments matched the request.
Compare accepted work under the same settings
Keep the same tool definitions and conversation history when comparing configurations. Check the action’s actual result and any clarification the assistant should have requested.
Report latency alongside accepted outcomes and memory use. Changing the reasoning mode or adding a drafter can change several of those quantities at once.
The repository was created before the October 23 release. Use the provider’s stated release date when describing availability, rather than treating repository creation as public launch.
Our Nemotron 3.5 Lightning article covers a separate, much larger checkpoint. The Nemotron retrieval guide examines models that perform other parts of an AI workflow.
Sources
- NVIDIA release and workload-specific measurements
- Model card, license and deployment variants
- Exact artifact size and repository metadata
What must remain loaded beside the language model in your intended local assistant?
Checked October 25, 2026. Prepared with AI-assisted research and drafting, with claims checked against the linked sources. MustHave.ai has not performed the proposed product tests.
