The number that does the work
Bolt's headline is 74B. The operating facts are 9B and 2 bits.
The mixture-of-experts architecture contains 74 billion parameters in total but activates approximately 9 billion for computation per token, according to NVIDIA. Different tokens can use different experts; the remaining parameters still need to be stored. NVIDIA says 2-bit quantization, combined with quantization-aware training, reduces the model's footprint to 24GB.
The more precise description is a 74B-total, approximately 9B-active-per-token agent model with a claimed 24GB compressed footprint. Actual deployment memory also depends on context length, KV cache, runtime overhead and the agent software running alongside it. NVIDIA has not yet published independent results establishing usable speed or quality on a 24GB machine. It is built to stay resident on a personal machine and handle routine work locally: coding help, monitoring, tool calls and background tasks through agents such as Hermes and OpenClaw. Whether 2-bit storage costs it accuracy on the tasks that matter is exactly the question that cannot be answered from a launch post. NVIDIA's own framing, faster than common dense models and stronger than similar-footprint MoE models, is a vendor comparison with no released weights behind it until 15 October.
Why NVIDIA is giving the model away
The open terms are part of the hardware pitch. Bolt is listed as fully open, including training data, weights and recipes, which lets developers fine-tune it and lets every runtime that matters for local inference support it on day one: llama.cpp, vLLM, Ollama, Unsloth and LM Studio. A closed model of the same size would still sell RTX Spark laptops. An open one gives the whole local-agent ecosystem a reason to optimise for them first.
PAIR points in the same direction. It left beta on 13 October as a free, open-source router that finds compatible PCs on a local network and distributes AI workloads across them, now with llama.cpp support and direct model downloads from Hugging Face. One machine is the product NVIDIA sells. A house or studio full of machines quietly acting as one inference pool is the habit it would prefer you to form, using its software as the layer in between.
NVIDIA says local LLM downloads are already five times higher than the total recorded last year, but provides no dataset, measurement period, download definition or methodology. The figure indicates what NVIDIA claims is rapid growth; its magnitude cannot be independently assessed from the announcement. The direction is plausible enough: local models have become good enough for agent plumbing, and routing easy work locally while reserving cloud models for harder queries is the setup agent apps in this same announcement, through Switchyard and similar routers, are being built to support.
The video and creator items buried in the same post
The announcement most relevant to video work is not the model. ComfyUI is getting RTX Video Frame Generation and RTX Video HDR as native nodes, alongside EXR export. NVIDIA's example starts with an LTX 2.5 clip at 1080p and 30 frames per second in SDR, then produces a smoother 60fps HDR version that can be graded properly. That is a practical finishing step for locally generated video: generate at a manageable rate and range, then use dedicated hardware features for the polish, instead of paying generation credits for the final frame rate.
Adobe's part needs its timing stated plainly. RTX Spark optimizations for selected Photoshop and Premiere AI features are available in beta, while Substance 3D updates are still forthcoming. NVIDIA and Adobe claim up to 2x improvement across selected creative workloads, not a universal doubling of application performance. Broadcast 2.3, with an offline file processor and agent controls through MCP, is available now. NVIDIA says a new DLSS 5 AI model is twice as fast internally and can increase frame rates by up to 1.4x in NBA 2K27 from 16 October; the model is scheduled for Hogwarts Legacy on 5 November and Where Winds Meet in late November.
What to test on 15 October
Three checks will settle most of the argument once the weights are public. First, task quality at 2 bits against the same prompts run on a less compressed MoE of similar active size, especially on tool use and instruction following, where small errors compound across an agent run. Second, real memory and speed on a 24GB machine with an agent harness attached, not a bare chat benchmark. Third, the licence text as actually published, because fully open was the promise in the blog and the document on release day is what a commercial user has to rely on.
If Bolt passes those, NVIDIA will have done something useful: a local agent model that is open enough to build on and small enough to leave running. If the quantization shows, the launch will still be instructive. It will have marked the point where the marketing number for a laptop model reached 74B while the model doing the work stayed in single digits.
Sources
- [1] NVIDIA Blog, "All Systems Go: Global PC App Ecosystem Ready for NVIDIA RTX Spark", 13 October 2026 (primary source for Bolt 74B specs, dates, PAIR, ComfyUI nodes, Adobe timing, DLSS 5 dates and the local-download claim)Read source
- [2] Blockchain.News, "NVIDIA Announces Software Ecosystem Updates Ahead of RTX Spark Laptop Launch", 13 October 2026 (secondary summary of the same NVIDIA announcement)Read source
- [3] Wccftech, "NVIDIA RTX Spark Gets 100s of Games & Anti-Cheat Support Ahead of Launch", 13 October 2026 (secondary coverage of the gaming and creative sections)Read source
- [4] Gadget Pilipinas, "RTX Spark Launch Week: Nemotron 3.5 Bolt and DLSS 5 Updates", 13 October 2026 (secondary coverage; notes Llama for Windows and NVIDIA MCP servers in the same rollout)Read source