NVIDIA said on October 28, 2026 that RTX Spark Windows laptops are available from ASUS, Dell, HP, Lenovo, Microsoft and MSI, with prices starting at $2,599. The same announcement introduces Nemotron 3.5 Transcribe, an open-weight model for working with recorded speech, and a control in NVIDIA App for sharing GPU resources between foreground applications and background AI agents.
The three changes fit together. A local agent can keep processing documents or audio while its owner edits video, writes code or plays a game on the same laptop. RTX Spark provides the shared hardware, and the NVIDIA App control lets the owner decide which workload gets priority. The important qualification is that model availability, optimized deployment options and laptop specifications are separate questions: a model being released does not mean every configuration will run it in the same way.
Six Laptop Makers, With Different Memory Configurations
RTX Spark combines an NVIDIA Blackwell RTX GPU and a Grace CPU in one superchip. NVIDIA specifies up to 128GB of unified memory and up to one petaflop of FP4 AI performance for the platform. Those are maximum specifications, not promises for every laptop or measures of performance in a particular app. FP4 also describes a low-precision AI calculation format; a peak FP4 figure is not a general-purpose speed rating.
The NVIDIA marketplace listings show why the $2,599 starting price needs context. The configurations visible there range from a Microsoft Surface Laptop Ultra with 24GB of unified memory to models with 128GB. One ASUS ProArt P16 with 128GB of memory and 2TB of storage is listed at $6,999.99. NVIDIA’s announcement gives the platform’s starting price but does not tie that price to a specific configuration in the cited material.
| Marketplace example | CPU and GPU cores listed | Unified memory | Storage listed |
|---|---|---|---|
| Microsoft Surface Laptop Ultra, EP 2 58079 | 18 CPU; 5,120 GPU | 24GB | 512GB |
| ASUS ProArt P16, 32GB configuration | 20 CPU; 6,144 GPU | 32GB | 1TB |
| Lenovo Yoga 9n 2-in-1, 16-inch | 20 CPU; 6,144 GPU | 64GB | 1TB |
| ASUS ProArt P16 H7607BA-XP99T | 20 CPU; 6,144 GPU | 128GB | 2TB |
Examples are from NVIDIA’s marketplace; they are not a complete price list.
Memory deserves more attention than the platform’s maximum AI-performance figure if the aim is to run a sizeable local model alongside normal Windows work. The CPU, GPU, model and foreground apps all draw on the system’s unified memory. A 24GB configuration and a 128GB configuration therefore present very different limits, even though both carry the RTX Spark name. My view is that buyers planning persistent agents should pick a model and workload first, then check the laptop’s actual memory and storage, instead of treating the $2,599 entry price as representative of the whole range.
NVIDIA App Can Prioritize Apps or Background Agents
The new resource-management control is in NVIDIA App > System > Advanced. According to NVIDIA’s announcement, it lets an RTX Spark user favor the foreground app for performance or give a background agent more GPU resources so its work progresses faster. Hardware-accelerated GPU scheduling must be enabled in Windows Settings for the feature to work.
That prerequisite is worth checking before relying on the control. An agent processing a long recording and a creative app rendering a preview can both demand GPU compute. Favoring the app is the sensible choice when responsiveness matters immediately; favoring the agent makes more sense when finishing its batch task matters more. The control gives users a way to express that preference without stopping the background task.
NVIDIA does not specify a fixed GPU percentage, a guaranteed response time or a per-agent quota in the announcement. It describes resource sharing, so users should not read the feature as a promise of isolated GPU capacity. It also will not add memory to a laptop that cannot fit the intended model and other open applications. Those are distinct constraints: the app control addresses competition for GPU work, while the configuration determines how much shared memory is available in the first place.
Nemotron 3.5 Transcribe Handles More Than Dictation
Nemotron 3.5 Transcribe is an open-weight, promptable speech model. NVIDIA says it can transcribe audio in 40 languages, including seven Indic languages, and handle speaker attribution, translation and questions about a recording. A developer could prompt it to “Transcribe this recording and label each speaker,” then ask for decisions made in the same meeting. Prompts can also supply product terminology or formatting conventions.
That combination has a practical appeal for a local workflow. A meeting assistant needs more than a stream of words: it may need to distinguish speakers, translate a passage and extract an action item. NVIDIA says one model can handle those tasks, which could simplify an application that would otherwise connect several speech and language models. Whether its answers are reliable enough for a particular recording still depends on the audio and the task; the announcement is not a substitute for checking an important transcript or decision summary.
NVIDIA reports approximately 4.3 billion active parameters and says the model ranks No. 1 on Hugging Face Open ASR and No. 2 among open-weight models on Artificial Analysis. Those are vendor-reported rankings, not measurements made for this article or evidence of how quickly a given laptop configuration will process audio. Benchmark position also says little by itself about a user’s microphone quality, terminology or mix of languages.
The release timing needs a careful reading. NVIDIA says the model is available on Hugging Face and describes it as available on RTX Spark, but also says quantized versions are coming soon for local deployment on RTX Spark and other NVIDIA systems. It names NeMo-Speech.cpp and vLLM as deployment routes for forthcoming NVFP4 versions. Quantization uses lower-precision representations to reduce a model’s resource demands; until those versions arrive, readers should not assume the planned NVFP4 deployment path is already available. NVIDIA has not provided, in the cited announcement, a configuration-by-configuration memory requirement or laptop transcription-speed result.
Open weights give developers a choice about where to process audio, including locally. Keeping a recording on a laptop can be useful when uploading it is undesirable, but that outcome depends on how the surrounding application is configured. A local model does not, by itself, prevent an agent or another connected service from sending data elsewhere.
LFM3 Broadens the Local-Agent Options
NVIDIA’s update also points to Liquid AI’s LFM3 family, which spans text, vision, audio, multimodal and decision models. NVIDIA says developers can run LFM3 models on RTX Spark through and , and can fine-tune them locally. It cites agents built with Hermes Agent, Pi and OpenClaw as possible users of these models.
Frequently Asked Questions
3 questions
1Which companies sell RTX Spark laptops?
NVIDIA names ASUS, Dell, HP, Lenovo, Microsoft and MSI. It says prices start at $2,599, while specifications and prices vary by configuration.
