NVIDIA Unveils Nemotron 3.5 Bolt, an Open 74B Model for Local Agents

A 74B MoE Squeezed Into 24GB

NVIDIA announced Nemotron 3.5 Bolt 74B-A9B on October 13, 2026. It is a mixture-of-experts model with 74 billion total parameters and roughly 9 billion active per token.

The headline number is the memory footprint. NVIDIA claims that 2-bit quantization combined with quantization-aware training shrinks the model to 24GB. That is NVIDIA’s claim, not an independently verified result. It matters because 24GB is the kind of memory budget a high-end consumer machine can actually carry.

The model is purpose-built for agentic AI on personal machines. That framing matters. Most flagship models still assume datacenter GPUs or cloud APIs. Nemotron 3.5 Bolt is aimed at the opposite end: a user running agents on their own hardware.

Weights release on October 15, one day before RTX Spark laptops go on sale on October 16. The timing is deliberate. The model and the machines are meant to arrive together.

Fully Open, Down to the Training Data

NVIDIA says the release will be fully open source. That includes the training data, the weights, and the recipes.

That combination is unusual. Many “open weight” releases stop at the weights. Including the training data and recipes gives researchers and builders a much clearer picture of what the model actually contains and how it was built.

At launch, the model will be available through llama.cpp, vLLM, Ollama, Unsloth, and LM Studio. Those are the standard distribution channels for locally run models. It will also be usable through agents such as Hermes and OpenClaw.

This breadth matters for adoption. A model is only as useful locally as the tools that can run it. Day-one support across the main runtimes removes the usual weeks of community porting work.

PAIR Leaves Beta

On the same day, NVIDIA moved PAIR — Personal AI Router — out of beta. PAIR is a free and open-source tool that distributes AI workloads across compatible PCs on a local network.

The update adds three things: support for llama.cpp, direct model downloads from Hugging Face, and compatibility with RTX Spark machines. In practice, that means a household or small office with several capable PCs could split model workloads between them instead of cramming everything onto one machine.

PAIR is a companion story to Nemotron 3.5 Bolt. One is the model. The other is the plumbing that makes running it on ordinary hardware more practical. A model that fits a 24GB footprint still benefits from a network of machines when the task is large or several agents run at once. PAIR gives that network a way to share the load without sending data to a cloud provider.

The Hugging Face integration matters for the same reason. Direct model downloads remove a manual step between reading about a model and running it. When the weights land on October 15, a PAIR user with compatible machines could pull the model and spread it across their local network without hunting for mirrors or converting formats.

The 2-Bit Question

The technical claim at the center of this announcement is aggressive. Two bits per parameter is far below the precision most local models ship at. NVIDIA pairs it with quantization-aware training, which teaches the model to tolerate the compression during training rather than applying it after the fact.

That approach is the right one if you want small files and usable quality. Whether it delivers both here is unknown. The accuracy trade-offs are unknown until the weights are out. Nothing NVIDIA has said so far changes that. The training method is a reason for optimism, not a substitute for measurement.

This is the pattern to watch with compression claims in general. A claimed footprint is a claim about file size and memory use. It is not a claim about quality. The two get conflated easily in launch coverage. NVIDIA’s 24GB figure describes the model’s size, not its performance. Performance is the part nobody outside NVIDIA has tested.

What Remains Unproven

The honest part of this story is short. NVIDIA has not published independent speed or quality results for the model running on 24GB machines. The accuracy trade-offs of 2-bit quantization are unknown until the weights are actually out and people can test them.

Vendor claims about compression are worth treating as claims. Quantization-aware training should preserve more quality than post-training quantization, and that is the approach NVIDIA says it used. Whether the result holds up on real agentic tasks will be settled after October 15, when the weights are public and independent testing becomes possible.

The timing is tight by design. Weights on October 15, RTX Spark laptops on sale October 16. NVIDIA wants the model and the hardware judged as a pair. Judging them as a pair means waiting for both to be in users’ hands.

The open release, down to training data and recipes, is the part that stands regardless of benchmarks. That part NVIDIA has committed to in writing. The rest is a testable claim with a test date: October 15.