What happened?
NVIDIA launched the Nemotron 3.5 Bolt 74B-A9B 24GB, a 74-billion-parameter language model (with 9 billion active per token) optimized for local AI agents on RTX Spark systems. The model uses 2-bit quantization and fits in 24 GB of memory.
Available from October 15, Nemotron 3.5 Bolt can be run via llama.cpp, vLLM, Ollama, LM Studio and Unsloth. It is designed for agentic tasks such as coding, research and tool use, balancing intelligence and speed.
Why is this important?
With the launch of RTX Spark laptops on October 16 — offering up to 128 GB of unified memory and up to 1 petaflop of AI — NVIDIA is pushing large model execution to local devices, reducing cloud dependency.
The model positions better than similarly sized alternatives in agent, web browsing and terminal use benchmarks, according to the company.
What changes in practice?
Developers and RTX Spark users (and compatible GeForce GPUs) can run personal agents locally with good performance. A smaller model, Bolt E4B 4GB, arrives on October 23 for game characters and conversational agents.
This accelerates on-device AI adoption on Windows Arm PCs, with gaming and creative tools also optimized for the platform.
By GeekikiBot
