NVIDIA used an October 13 post to line up the software for its RTX Spark laptops, which go on sale Friday, October 16, and to announce Nemotron 3.5 Bolt 74B-A9B, a mixture-of-experts model that NVIDIA says fits in 24GB through 2-bit quantization-aware training and arrives Thursday, October 15. NVIDIA’s October 13 post Taken together, the question for local AI moves from which hardware can run a 70B-class model to whether a 2-bit 70B-class model is good enough to use. Neither has shipped, so this piece sticks to what NVIDIA and its partners have published and what to check once both are out.
Update, October 15, 2026: NVIDIA released Bolt on Hugging Face today, with model cards, a technical report and a research page, so most of what this piece called unpublished now has an answer. The license is OpenMDW-1.1, the context window is up to 256K tokens, and the 24GB figure is a memory budget for text-only use at a 128K context rather than a file size. NVIDIA also published benchmarks, including a comparison with the unquantized model. NVIDIA’s October 15 post Bolt GGUF model card Bolt technical report Dated notes in the sections below give the details. The laptops were still on preorder when we checked this evening, the day before their on-sale date.
What NVIDIA announced
The laptops come first. NVIDIA describes RTX Spark as offering “up to 128GB of unified memory and 1 petaflop of AI compute”. NVIDIA’s October 13 post That October 13 post does not give a precision for the petaflop, but the October 7 launch post and the product page both say it is FP4. NVIDIA’s October 7 post RTX Spark product page FP4 is a low-precision format, and the figure should not be compared with FP16 or BF16 throughput on another machine.
The October 7 post, from a Microsoft event in San Francisco, named Acer, ASUS, Dell, HP, Lenovo, Microsoft, MSI and Gigabyte as laptop makers, said preorders had opened, and put compact RTX Spark desktops on sale in November. NVIDIA’s October 7 post NVIDIA lists no prices. Microsoft’s store page for the Surface Laptop Ultra shows $2,599.99 next to a preorder button, without saying which memory and storage configuration that buys. Surface Laptop Ultra
Update, October 15, 2026: NVIDIA’s product page still says “Preorder Now” and still lists no prices. It names six launch laptops: ASUS ProArt P16, Dell XPS 16, HP OmniBook Ultra 16, Lenovo Pro 9n, Microsoft Surface Laptop Ultra and MSI Prestige N16 Flip AI. RTX Spark product page The OEM prices we could find are starting prices only. HP expects to sell the OmniBook Ultra 16 on HP.com from October 16 starting at $3,199.99, and the OmniBook X 14 from $2,999.99. HP, October 7 Dell says the XPS 16 Creator Edition is on preorder at Best Buy from $3,799.99, with Dell.com to follow. Dell, October 7 Microsoft’s $2,599.99 Surface Laptop Ultra listing still names no configuration. Lenovo said in September that pricing would be announced later, ASUS’s ProArt P16 spec sheet has no price, and MSI’s US store blocked our requests. Lenovo, September 3 ASUS ProArt P16 specifications None of these prices says how much memory it buys.
| Item | Date | Source |
|---|---|---|
| RTX Spark laptops on sale | Fri, Oct 16 | NVIDIA |
| Nemotron 3.5 Bolt 74B-A9B (24GB) | Thu, Oct 15 | NVIDIA |
| Nemotron 3.5 Bolt E4B (4GB), for NVIDIA ACE game characters | Fri, Oct 23 | NVIDIA |
| RTX Spark compact desktops | November | NVIDIA |
| GTC Berlin keynote | Wed, Oct 21 | NVIDIA |
128GB is a ceiling
The product page lists two RTX Spark configurations: up to 128GB of LPDDR5X with a 6,144-core Blackwell RTX GPU and a 20-core Grace CPU, and up to 64GB with a 5,120-core GPU and an 18-core CPU. RTX Spark product page The page does not say which of those ship in laptops and which in desktops. Microsoft’s footnote says memory “varies by configuration”. Surface Laptop Ultra
If you are buying one of these to run large models, check the memory in the exact configuration you order. With unified memory, the operating system, the applications you keep open and the model’s KV cache all draw from the same pool as the weights. A 64GB machine and a 128GB machine allow very different model choices, and the launch materials don’t say how much of either is available to the GPU. Our Local LLM Machines guide covers the arithmetic for weights plus context.
Update, October 15, 2026: Memory goes lower than the two configurations above suggest. The ASUS ProArt P16 spec sheet lists 24, 32, 64 and 128GB of LPDDR5X across the two chip variants, without saying which ship in the US. ASUS ProArt P16 specifications NVIDIA’s buying guide sorts RTX Spark machines into 24, 32, 48, 64 and 128GB tiers and says agents run in the cloud on the two smallest. RTX Spark: Help Me Choose NVIDIA’s own llama.cpp recipes for Bolt, covered below, start at the 64GB RTX Spark.
Bolt: 74 billion parameters at 2 bits
NVIDIA calls Bolt “a 74B parameter, 9B-active-parameter mixture-of-experts (MoE) model, packed into 24GB using 2-bit quantization with advanced quantization-aware training techniques to retain quality.” It lists llama.cpp, vLLM, Ollama, Unsloth and LM Studio as launch runtimes, with access also through the Hermes and OpenClaw agents. NVIDIA’s October 13 post
At a flat 2 bits, 74 billion parameters come to about 18.5GB, so the published size suggests some layers are kept at higher precision. That is our arithmetic; NVIDIA has not described the quantization layout. The post also doesn’t say whether 24GB covers anything beyond the weights. If it is weights only, a 24GB desktop GPU would have almost nothing left for the KV cache, and a usable context window would need offloading or a smaller cache. Our explainer on the KV-cache memory wall shows how quickly that cache grows with context length. Bolt’s context length and architecture are not in the announcement either. Nemotron 3.5 Lightning, an earlier model in the same family, used a hybrid Mamba-2 and MoE design with few attention layers, a design that generally needs a smaller KV cache than full attention, but nothing published says Bolt does the same. Our Lightning coverage
Update, October 15, 2026: The model card and technical report now give the architecture, context length and quantization layout. Bolt is a compressed version of Nemotron 3.5 Super VL, built as a Mamba-2 and Transformer hybrid with latent MoE layers and a C-RADIO v4-H vision encoder, and it takes images and video as well as text. Of its 88 layers, 40 are Mamba-2, 40 are MoE and 8 are attention, and each token goes to 9 experts, down from 22 in the parent model. The context window is up to 256K tokens. Bolt BF16 model card config.json As our arithmetic suggested, the precision is mixed. The routed and shared experts are stored in IQ2_XXS at 2.0625 bits per weight, the Mamba projections in NVFP4, and attention, the latent MoE projections, the embeddings and the output layer in Q8_0. Bolt technical report
The 24GB figure does cover more than the weights. NVIDIA’s estimate at batch size 1 and a 131,072-token context puts the weights at 22.01GB, the KV cache, Mamba state and activations at 0.87GB and a speculative-decoding drafter at 0.96GB, for 23.83GB in total. The vision encoder and projector, another 1.68GB, are excluded, so the figure covers text-only use. Bolt technical report The KV cache accounts for only 0.54GB of the 0.87GB, because only the 8 attention layers keep one, each with 2 KV heads. Even so, NVIDIA’s llama.cpp recipes don’t target a 24GB card. The GGUF card gives a 28GB recipe with a 65,536-token context for a 64GB RTX Spark or a discrete GPU with 32GB or more, and a 32GB recipe with the full 262,144 tokens for a 128GB RTX Spark or a GPU with 48GB or more. Bolt GGUF model card Ollama lists its build at 25GB. Ollama library
On quality, NVIDIA says Bolt is “faster than most popular dense models while smarter than most popular MoE models with a similar memory footprint.” NVIDIA’s October 13 post That is a claim without numbers. The post has no benchmarks, no comparison against a higher-precision Bolt, and no figure for what 2-bit quantization costs. The 9B active parameters should make generation fast relative to the model’s size, but NVIDIA has not shown how much reasoning survives at 2 bits, and that will take independent testing.
Update, October 15, 2026: NVIDIA has now published numbers. The 24GB model card compares 2-bit Bolt with Qwen3.6-35B-A3B, Gemma 4 26B-A4B and Nemotron 3.5 Lightning 30B-A3B, all at BF16, on 13 benchmarks, and Bolt has the top score on 10 of them, including 86.43 on GPQA Diamond and 66.75 on BrowseComp. Bolt 24GB model card The BF16 checkpoint’s card adds the comparison we asked for. The unquantized model scores higher on 10 of the 13 benchmarks, by 2.4 to 4.9 points, ties on one and scores lower on two; GPQA Diamond drops from 88.83 at BF16 to 86.43 at 2 bits. Bolt BF16 model card The technical report calls the quantization “very aggressive and lossy by design” and says the aim is the best accuracy for the size, not accuracy close to BF16. Bolt technical report Beyond a footnote saying NVIDIA measured Qwen3.6 at 51.2 on MCP-Atlas against a reported 62.8, the cards don’t describe how the benchmarks were run. These are vendor-run comparisons. NVIDIA’s release post calls Bolt “#1 open model in its class on PinchBench,” while its own table has Qwen3.6 slightly ahead, 88.07 to 87.82. NVIDIA’s October 15 post
NVIDIA’s research page also gives speeds, measured with llama.cpp on a 128GB Surface Laptop Ultra. On a coding throughput test, Bolt with its drafter ran at 83.99 tokens per second, against 96.63 for Qwen3.6-35B-A3B, 100.56 for Gemma 4 26B-A4B and 28.26 for the dense Qwen3.8 27B, all three running 4-bit builds. NVIDIA research: Nemotron 3.5 Bolt By NVIDIA’s own figures, Bolt is slower than the two smaller MoE models and about three times as fast as the dense one. Nobody has published independent results yet.
License: claimed open, not yet published
NVIDIA writes that “like other Nemotron models, Bolt 74B is fully open source — including training data, weights and recipes.” NVIDIA’s October 13 post No license is named, and as of October 14 there is no model card to check. Lightning shipped under OpenMDW-1.1, a permissive license, as our August coverage noted, so Bolt may follow, but that is a guess. Read the model card and the license before Bolt goes into a commercial product, and confirm that the training data has actually been released rather than just described.
Update, October 15, 2026: The model cards name the license: the OpenMDW License Agreement, version 1.1, as with Lightning. It grants rights to use the model “without restriction,” places no restrictions on outputs, and asks anyone redistributing the model to keep a copy of the agreement and the notices. OpenMDW-1.1 The card says the model “is ready for commercial or non-commercial use.” Bolt BF16 model card The LICENSE files don’t carry the text, though: the BF16 repository’s file is empty and the 24GB and GGUF repositories have none, so the terms come from the card and its link. BF16 repository files
The claim of fully open training data and recipes holds for the Nemotron family more than for Bolt itself. The card lists the Nemotron pre-training and post-training datasets, which NVIDIA publishes on Hugging Face, with several of the pre-training sets behind a manual access request. Nemotron pre-training datasets The technical report says Bolt’s supervised fine-tuning used the same data blends as Nemotron 3.5 Super VL. Bolt technical report We found no dataset published for Bolt specifically: the two dataset IDs in the card’s metadata don’t resolve to public datasets. NVIDIA’s Nemotron recipe repository on GitHub has no Bolt recipe either. NVIDIA-NeMo/Nemotron
Runtime news that applies to other models too
The same post says decision models are now supported in llama.cpp and Unsloth on RTX Spark and GeForce RTX GPUs, that llama.cpp released Llama Windows, a Windows-native companion app, on October 13, and that Microsoft is adding llama.cpp as a backend to Windows ML, with no date given. NVIDIA also says decision models run “up to 200x faster than traditional agents” and gives no method for that figure. NVIDIA’s October 13 post For anyone following open decision models, the llama.cpp support is the useful part; the multiplier can’t be checked.
What to test first
Bolt is the easier one to evaluate, because it runs on hardware many teams already have. Once it appears in llama.cpp or Ollama:
- Load it on a 24GB GPU and record the longest context that fits without offloading, along with tokens per second at that length.
- Run your own task set against the higher-precision model you use now, and count corrections rather than reading off a leaderboard.
- Check the model card for license, context length and quantization format before anything else depends on it.
Update, October 15, 2026: Bolt is in llama.cpp now. Support was merged today and shipped in build b11669 and the v0.7.0 release. llama.cpp pull request 30520 llama.cpp v0.7.0 Ollama has a nemotron-3.5-bolt entry at 25GB with a 256K context and says Apple Silicon and Metal are not currently supported. Ollama library Unsloth has published a GGUF build and a guide. Unsloth GGUF NVIDIA’s vLLM instructions call for a nightly build; the latest tagged vLLM release came out on October 5. Bolt 24GB model card LM Studio’s model catalog had no Bolt entry when we checked, although an lmstudio-community GGUF is on Hugging Face. LM Studio models For the first test above, NVIDIA’s smallest llama.cpp recipe needs 28GB, so a 24GB card is outside what NVIDIA has documented. On a 64GB laptop, that recipe uses a 65,536-token context.
For the laptops, wait for configurations with prices attached. A $2,599.99 listing with no memory spec cannot be compared with a desktop GPU build, and the 128GB in the headline describes the top configuration, not the one a given SKU ships with.
If Bolt holds up on real work at 2 bits, 24GB cards gain a 70B-class option and a 64GB laptop could hold it alongside a longer context. If it doesn’t, what lasts from this announcement is the llama.cpp work, which helps every other model too.