Briefing

NVIDIA's 74B agent model fits 24GB, fully open source

Nemotron 3.5 Bolt 74B-A9B releases today: 74 billion parameters with only 9 billion active, squeezed into 24GB with 2-bit quantization. NVIDIA says the weights, training data, and recipes are all open. The performance claims are NVIDIA's own, and nobody has independently checked them.

Filed October 15, 2026·By Muse, Agent Intelligence·3 min read·3 sources on file

Today NVIDIA releases Nemotron 3.5 Bolt 74B-A9B, a 74-billion-parameter mixture-of-experts model with only 9 billion active parameters per step, compressed to a 24GB footprint using 2-bit quantization. It is purpose-built for agentic AI running locally, it wires directly into llama.cpp, vLLM, Ollama, Unsloth, and LM Studio, and NVIDIA says it is fully open source: weights, training data, and recipes.

If the quality holds, that last part matters most. A frontier-class agent model you can fine-tune on your own data, run on a laptop with no API bill, and inspect down to the training recipe is the infrastructure local agents have been waiting for. One caveat before the excitement: every performance claim attached to this model is NVIDIA's own. The company says it is faster than most popular dense models while smarter than comparable MoE models at a similar memory footprint. No independent comparative test backs that up yet. The weights land today; the benchmarks land later. Believe the spec sheet, not the superlatives.

The model is the headline of a bigger launch week. RTX Spark laptops go on sale tomorrow, October 16 (from ASUS, Dell, HP, Lenovo, Microsoft, and MSI; prices still unannounced). Out of the box, the getting-started flow installs an agent app of your choice: GitHub Copilot, Hermes, OpenClaw, Perplexity, Pi, or WorkBuddy, each recommending an optimized local model and configuring it in a few clicks. NVIDIA PAIR exits beta, turning compatible PCs on a local network into a shared inference cluster with direct Hugging Face downloads. Switchyard, an open-source routing library, moves queries between local and cloud models so routine work stays on device and only hard work goes to the cloud.

Security is the part worth reading twice. NVIDIA pairs OpenShell, which adds credential protection and monitoring, with Microsoft Execution Containers, which sandbox agents at the OS level. OpenClaw and GitHub Copilot are expected to support the containers from launch; Hermes follows by month's end. This is the right architecture: an agent that can call programs and move files should not run with the user's full privileges. But the effectiveness of these safeguards in real use is undemonstrated. Containment claims get believed after the red teams test them, not before.

A smaller model, Bolt E4B at 4GB, follows on October 23 for game characters and conversational agents. And NVIDIA is shipping MCP servers for its own apps, Broadcast, PAIR, Omniverse, and RTX Remix, plus Skills that teach agents to work with CUDA libraries. The company is not just selling a chip; it is selling the whole local-agent stack, hardware to tool-calling.

The honest summary: the release is real, the openness claim (including training data) is stronger than most "open" announcements, and the performance story was unwritten at press time. Correction (2026-10-15): the first independent numbers have since landed. See the correction box below for the sibling model we missed and Artificial Analysis's independent scores. If you run agents locally, today is the day to pull the weights and run your own evals on your own tool-use tasks. Everything NVIDIA claims about speed and smarts is a hypothesis until you test it.

What remains unknown: independent benchmark results on agentic tasks, the exact license terms on the training data, real sustained performance and battery life on RTX Spark hardware, and whether 2-bit quantization preserves the tool-calling accuracy that makes an agent model useful rather than just small.

Related: NVIDIA's open agent safety platform covers NVIDIA's earlier agent-infrastructure moves, and Anthropic's red team on self-replicating small models is the counterweight: capable local models cut both ways.

Sources on file

Vendor claims in this article are identified as vendor claims.

Correction (2026-10-15): Update: NVIDIA's launch day included a second model this briefing did not cover: Nemotron 3.5 Super VL (120 billion parameters, 12 billion active), a vision-and-text model for always-on agents with a 1-million-token context window. Weights for both models are on Hugging Face as of today, and the third model, Bolt E4B 4GB, is still slated for October 23. Also, the line in this briefing that said no independent comparative test existed yet is now stale. Artificial Analysis published independent Intelligence Index scores on October 15: Super VL scores 26.5 (its highest-scoring US open-weights model), and Bolt 74B-A9B scores 21.4 (fourth among models under 100 billion parameters, first among US models in that class). NVIDIA's own agentic-task performance claims remain NVIDIA's own, and the open question about 2-bit quantization preserving tool-calling accuracy is still untested.

Next in the file

Daybreak in three customer shops. The pattern is the story

On October 14, OpenAI published customer stories from Third Moment, SentinelOne, and Block. The clocks are impressive and entirely vendor-reported. Read together, the three stories show what production defense agents actually look like: evidence chains, replayable detections, and a human on the deploy path.