NVIDIA announced on Friday, October 23, 2026, the availability of Nemotron 3.5 Bolt E4B, a small open language model designed for conversational agents that need to run locally with low memory use and fast responses.
What happened?
Bolt E4B is built on top of Gemma 4 E4B with Nemotron-specific post-training. According to NVIDIA’s official page, the model combines natural conversation with multi-turn tool use. It fits in approximately 4GB for target workloads and is now available on Hugging Face, with open weights, data and training recipes.
NVIDIA highlights uses such as voice and device assistants, interactive characters (game NPCs, for example) and embedded agents in robotics and IoT. In one official example, the model runs on a Jetson Orin Nano 8GB alongside Cosmos for vision, controlling a Reachy Mini robot.
Why it matters
Frontier models typically require data centers. Bolt E4B points in another direction: agentic intelligence that fits on RTX PCs, Jetson boards and edge systems. For the geek audience, this means potential for NPCs that actually converse and act locally, offline assistants and more responsive robots without always depending on the cloud.
NVIDIA states that the model delivers roughly 40% higher average quality on conversational agent benchmarks compared to the previous Nemotron 3 Nano 4B and generates up to 80% fewer tokens in reasoning-enabled comparisons. Context is 64k tokens and it supports 35 languages.
In practice
The model is optimized for NVIDIA hardware (GeForce RTX, Jetson, DGX Spark). It is not meant to compete on general coding leaderboards; the goal is efficiency and suitability for real-time interaction with people and tools.
Sources
Transparency: This content was created, edited or reviewed with the aid of artificial intelligence. Information was cross-checked with public posts on X and sources available on the internet. Consult the original sources for the full context.
Por GeekikiBot
