October has been the month of the closed model. Meta shipped Muse Wonder behind an API. OpenAI shipped GPT-6 behind ChatGPT. Google shipped Gemini 4 Argon behind its API and app. Every one of them: rent intelligence by the token, data leaves your machine.
Then NVIDIA went the other way. On October 13 it announced Nemotron 3.5 Bolt 74B-A9B — a 74-billion-parameter agent model it says fits in 24GB of memory, with the weights, the training data, and the training recipes all published. Not a demo. Not a gated preview. The whole thing, yours to download.
It was scheduled to land today, October 15, one day before the RTX Spark laptops it was built for go on sale. One honest note up front: as of this writing the weights were not yet confirmed live — everything so far is pre-launch coverage of a release scheduled for today. Check Hugging Face or Ollama for the live download rather than assuming it is already there.
What "74B in 24GB" actually means
The name is a spec sheet in miniature. 74B is the total parameter count. A9B is the roughly 9 billion parameters active on any given token. 24GB is the compressed footprint NVIDIA is claiming for the weights.
Bolt is a sparse mixture-of-experts model. Only about 9B parameters work on any one token, so the compute per token looks like a 9B model while the stored knowledge spans 74B. The catch: inactive parameters still have to be stored. Mixture-of-experts saves compute, not storage.
The storage trick is 2-bit quantization with quantization-aware training — parameters stored at 2 bits instead of the usual 16, with the model trained to expect the compression rather than squeezed after the fact. That is how 74 billion parameters land near 24GB.
And the fine print matters. 24GB is NVIDIA's claim about the weights, not a full system requirement. Context length, the KV cache, runtime overhead, and agent software all add memory. Treat 24GB as the floor, not a comfortable target.
Open vs closed: two theories of AI
This is where Bolt genuinely breaks from the month's other launches. Muse Wonder is closed weights behind an API at $4/$20 per million tokens. GPT-6 Sol is closed weights behind ChatGPT and OpenAI's APIs at $2/$10. Gemini 4 Argon is closed weights behind Google's API and app at $2/$10 intro pricing. Bolt is free to download — you pay for hardware and electricity.
The practical consequences are concrete. Open weights mean you can fine-tune the model for your own agents instead of prompt-engineering around someone else's API. You can run it entirely offline, which matters for privacy-sensitive work and anyone whose data cannot legally leave their machine. Researchers can inspect the training data and recipes, not just prompt a black box.
The honest counterweight: open does not mean equal. The closed models still lead on independent benchmarks, and Bolt has none yet. NVIDIA's "faster than dense models, smarter than same-footprint MoE models" line is NVIDIA grading its own homework. There are no independent benchmarks for Bolt 74B — do not treat that claim as measured fact.
How you actually run it
NVIDIA named five runtimes at launch: llama.cpp, vLLM, Ollama, Unsloth, and LM Studio, plus the Hermes and OpenClaw agent frameworks. The easy path is Ollama or LM Studio — pull the model by its official name and run it. The tinkerer path is llama.cpp (also the engine behind NVIDIA's new free Llama for Windows app). vLLM serves it as an API endpoint; Unsloth is for fine-tuning it on your own data.
Two warnings before you download anything. First, check the official listing for the exact model name — wrong or spoofed listings appear fast around popular releases, and guessing a model ID can fetch something that is not the official model. Second, verify the license before building a business on it. The weights were not yet public at announcement, so the exact terms are still to be confirmed when the release lands.
The ecosystem around it
Bolt does not arrive alone. It is the software flagship for RTX Spark laptops (up to 128GB unified memory, on sale October 16 — NVIDIA has announced no retail prices yet). And it arrives alongside PAIR, the free, open-source Personal AI Router that graduated from beta on October 13.
PAIR spreads inference jobs across the PCs on your home network. One precision point people get wrong: it routes whole jobs to one machine. It does not pool GPU memory. Three PCs with 24GB each do not become one 72GB virtual GPU. What you get is parallel throughput — NVIDIA's own demo cut a five-subagent inbox task from 18 minutes on one machine to under 9 minutes spread across three.
The verdict
Bolt is the most interesting AI release of October precisely because it refuses the month's dominant pattern. The proof is still pending: weights not confirmed live at write time, no independent benchmarks, an aggressive 2-bit quantization whose quality cost only real testing will reveal. Those are not reasons to dismiss it. They are reasons to test it instead of believing it.
If you want a capable agent model that runs entirely on your own hardware, with privacy, offline use, and fine-tuning — this is the one to watch. If you need maximum raw capability and are happy paying API prices, the frontier APIs remain the answer. And the hybrid setup — Bolt for the routine, a frontier API for the hardest 10% — is the most sensible default for production work right now.
I broke down the full comparison table, the run-it-yourself walkthrough, PAIR, RTX Spark, and every honest caveat on my site:
Full analysis: https://toolpilotpro.com/nvidia-nemotron-3-5-bolt-guide
More AI tool guides at toolpilotpro.com — the Gemini 4 Argon guide and the Muse Wonder comparison are good places to start.
Top comments (0)