From a scheduled date to visible weights
NVIDIA’s October 15 announcement scheduled Nemotron 3.5 Bolt E4B for October 23. Both official Hugging Face cards now list that release date, and their public file views list model weights. This is an availability follow-up to that plan; the exact time the repositories became public is unverified.
NVIDIA describes the model as further post-trained from Google’s Gemma 4 E4B-it for conversational and device assistants. The cards list eight billion total parameters, four billion effective, and text input and output. Vendor benchmark tables are separate from independent testing.
Choose the artifact before estimating the fit
The BF16 card describes a full-precision release aimed mainly at customization, post-training, and evaluation. Its file view lists a 15.9 GB model tensor file plus a 110 MB shared-KV tensor. The “4GB” sibling is the Q4_K_M quantized GGUF release: its named model file is listed as 6.18 GB, with the repository totaling 6.21 GB.
A download size is not the total memory needed to run a model. Runtime overhead, context, and the selected implementation also matter. The cards identify OpenMDW 1.1 governing terms and Apache 2.0 for the underlying Gemma 4 model; read the applicable terms and current serving instructions for the exact variant.
Our take
For developers exploring a local conversational assistant, the useful next step is a bounded test of the exact artifact on their own supported hardware. Check known answers, tool-call handling, incomplete outputs, latency, and total memory use before making it part of a business workflow.
A small-model label or a vendor comparison does not establish that a particular machine or task will work well. We inspected the cards and file listings; we did not download the weights, run the model, or verify minimum RAM or VRAM requirements.