NVIDIA Ships Nemotron 3.5 Bolt 74B for Local Agents: 9B Active Params in 24GB on RTX Spark

NVIDIA has launched Nemotron 3.5 Bolt 74B-A9B, a mixture-of-experts language model built specifically to run agentic AI workloads locally on its new RTX Spark platform. The model packs 74 billion total parameters with 9 billion active, compressed to a 24GB footprint through 2-bit quantization that was applied during quantization-aware training to preserve model quality rather than degrade it after the fact.

NVIDIA's positioning splits the difference between the two failure modes of local models. The company says Bolt is faster than most popular dense models at similar memory footprints while scoring higher than most popular MoE models in the same memory class — a speed-versus-intelligence balance aimed at agent loops, where latency compounds across every tool call. As with all vendor benchmarks, that characterization is NVIDIA's own and has not been independently verified.

The openness story is unusually complete for a frontier-adjacent release: training data, weights and training recipes are all published, letting developers fine-tune and rebuild rather than just consume. Day-zero runtime support spans llama.cpp, vLLM, Ollama, Unsloth and LM Studio, and the model is directly usable from agent harnesses including Hermes and OpenClaw — the same open-source agent runtimes that have been driving local-model demand all year.

Arriving alongside the model is the general-availability release of PAIR, NVIDIA's Personal AI Router, which exited beta the same day. PAIR is a free, open-source tool that discovers compatible PCs on a home network, installs models and inference engines on them, and spreads AI workloads across whatever machines are available — turning a household of devices into one pooled inference resource. The GA version adds llama.cpp support, direct Hugging Face model downloads, customizable inference engine settings and RTX Spark support, and runs on Windows, Linux and macOS. NVIDIA says the community contributed more than 10 features and fixes in the month since PAIR's first release.

The timing is no accident: RTX Spark laptops — built on the N1X superchip pairing a Grace CPU with Blackwell graphics and up to 128GB of unified memory — go on sale October 16, and NVIDIA plans to publish a hands-on AI playbook the same day covering first-agent setup, AI office assistants and development environments. Bolt 74B is effectively the launch model for that hardware story: a system that claims frontier-adjacent agent intelligence without sending a single token to the cloud.

The bet underneath all of this is that local inference becomes a real deployment tier rather than a hobbyist niche. If a 24GB quantized MoE can hold up on real agent tasks, the addressable market stops being developer toys and starts being every PC sold in the next upgrade cycle — which is precisely the market NVIDIA's data-center business cannot afford to cannibalize, and would rather own on both ends.