TL;DR
Today is the day NVIDIA’s Nemotron 3.5 Bolt 74B-A9B is supposed to land — the 74-billion-parameter agentic model squeezed into 24GB with 2-bit quantization, promised fully open source. As of this morning the weights were not on HuggingFace yet; they are expected later today through llama.cpp, vLLM, Ollama, Unsloth, and LM Studio. Tomorrow the RTX Spark laptops go on sale, and next week brings a smaller Bolt sibling built for game NPCs.
Release-day status: weights inbound, not here yet
October 15 was the date NVIDIA named for the Bolt 74B weights, and as of this morning they had not appeared: a direct query of HuggingFace’s model API for Nemotron 3.5 Bolt returned nothing, and an independent check by gurujicorp found the same. NVIDIA’s own schedule says today, so expect them later in the day rather than reading the delay as a slip.
When they land, the routes are the ones NVIDIA announced: llama.cpp, vLLM, Ollama, Unsloth, and LM Studio, plus direct access from agent harnesses like Hermes and OpenClaw. The ‘fully open source — weights, training data, and recipes’ promise is the thing to verify on arrival: the license file published with the weights is the document that actually matters, not the launch blog.
Tomorrow: RTX Spark laptops go on sale
Friday, October 16 is the on-sale date for RTX Spark laptops from Lenovo, Acer, and other NVIDIA partners, per the joint NVIDIA-Microsoft announcement from October 7. The machines pair up to 128GB of unified memory with a petaflop of AI compute, and NVIDIA says thousands of games, apps, and models will be ready at launch.
The software side is staged: agent apps including GitHub Copilot, Hermes, OpenClaw, Perplexity, Pi, and WorkBuddy get streamlined setup with security through Microsoft Execution Containers, and a first-run screen lets buyers pick an agent app that then recommends a local model and configures it. Adobe’s AI optimizations for Photoshop and Premiere are in beta with claimed gains up to 2x; Substance 3D improvements come later. No retail pricing has been published yet.
October 23: a smaller Bolt for game NPCs
Also new since our first story: NVIDIA announced a second model in the line, Nemotron 3.5 Bolt E4B — a 4GB compact model for fast dialogue and interactive game characters, built for NVIDIA ACE, the company’s in-game AI character toolkit. It launches October 23 alongside SkyrimNET, a platform for AI-driven NPCs in The Elder Scrolls V: Skyrim, where characters hold dynamic, contextual dialogue with the player.
If Bolt 74B is the ‘agent that lives on your laptop’ play, E4B is the latency-sensitive one: real-time responses where every millisecond counts. Two outlets corroborate the October 23 date and the SkyrimNET pairing.
Three things to check when the weights land
Once the files are downloadable, the vendor claims meet reality. The three checks that settle it: first, task quality at 2-bit quantization — run the same tool-use and instruction-following prompts against a less-compressed MoE of similar active size and see whether the compression shows, because small errors compound across an agent run. Second, real memory and speed on an actual 24GB machine with an agent harness attached, not a bare chat benchmark. Third, the license text as published — ‘fully open’ was the blog promise; the document on release day is what a commercial user has to rely on.
Our take has not changed since Tuesday: if Bolt passes those checks, NVIDIA has built something genuinely useful — a local agent model open enough to build on and small enough to leave running. If the quantization shows, the launch is still instructive: it marks the point where the marketing number for a laptop model hit 74B while the model doing the work stayed in single digits.