NVIDIA brings on-device game characters to RTX Spark

NVIDIA has introduced Nemotron 3.5 Bolt E4B through its In-Game Inferencing SDK, enabling developers to run AI game characters locally on RTX Spark. NVIDIA says the model matches the response quality of Gemma 4 26B while generating up to 80% fewer output tokens than comparable models. Those quality comparisons rely on NVIDIA’s evaluations and have not been independently reproduced.

The model supports conversation history, game API calls, configurable reasoning and dialogue across 35 languages. Local processing removes cloud network delays, while shorter responses are intended to make conversations more responsive. NVIDIA’s 4GB runtime activation memory figure does not represent the model’s total footprint: the downloadable GGUF file is approximately 6.18GB and requires a GPU with at least 6GB VRAM to load comfortably. Developers can test memory-aware characters through SkyrimNET, an open-source modding platform that now supports RTX Spark.

The developer update also adds RTX Spark support for all eight RTX Kit neural rendering SDKs and native Windows on Arm versions of Nsight profiling and debugging tools. These additions help studios build and optimize native Arm64 games. RTX Spark’s unified memory lets graphics and AI share resources without copying data between separate memory pools, but simultaneous rendering and inference still compete for compute and memory bandwidth. Broader adoption depends on studios shipping native games and integrating on-device characters into commercial releases.