Model release · NVIDIA
NVIDIA releases the open-weight Nemotron 3.5 Bolt E4B 4GB model
- Released
- Review updated
NVIDIA has released Nemotron 3.5 Bolt E4B 4GB, further post-trained from Gemma-4-E4B-it for local assistants, interactive characters and edge devices. The model card lists 8B total parameters, 4B effective parameters, Q4_K_M quantization and text-only input and output of up to 64K tokens each. In the official Reachy Mini demo, Bolt handles conversation and NVIDIA Cosmos handles vision; both run locally on Jetson Orin Nano 8GB. Vision is not a capability of Bolt itself. Weights are available under OpenMDW 1.1 for commercial and non-commercial use. Training data and recipes are provided, but some datasets require access approval. Scale Labs reports a 56.03% MultiChallenge score and first place among small models; this external result is separate from the model card’s internal evaluation.
Availability
Weights are available on Hugging Face under OpenMDW 1.1 for commercial and non-commercial use; some training data requires access approval.