RTX Spark laptops from six makers run AI agents on the device, starting at $2,599
Nvidia launched RTX Spark laptops on Friday with up to 128GB of shared memory and a 74-billion-parameter model it says fits in 24GB. A local agent pilot is now possible on Windows, though every figure so far comes from Nvidia.
The Product Desk
Key figures, each marked unconfirmed: Nvidia's Nemotron 3.5 Bolt model is squeezed into 24GB of memory. RTX Spark laptops have up to 128GB of shared memory. The model has 74 billion parameters.
| Figure | Value | Note | Claim |
|---|---|---|---|
| Memory Nemotron 3.5 Bolt is squeezed into | 24 GB | unconfirmed | 2 |
| Laptop shared memory | 128 GB, up to | unconfirmed | 8 |
| Model parameters | 74 billion | unconfirmed | 1 |
What happened
- Each laptop runs on one superchip that pairs a Blackwell RTX graphics chip with a Grace processor, rated at up to 1 petaflop of AI performance.
- Nemotron 3.5 Bolt, which Nvidia released on Thursday, activates 9 billion of its 74 billion parameters at a time.
- Nvidia's OpenShell runtime and Microsoft Execution Containers let users set what agents may do, and OpenClaw and GitHub Copilot already support the containers.
- ASUS, Dell, HP, Lenovo, Microsoft and MSI make the laptops, with prices starting at $2,599 and varying by region.
Why it matters
- What it limits constraint Hermes users wait until October for both Switchyard routing and OpenShell controls.
- What it costs cost A ten-seat pilot costs at least $25,990 in hardware, before anyone knows which memory tier the agents need.
- What's now possible capability A team can examine and test the model with tools it already uses, because the weights, training data and recipes are open and it works with runtimes such as Ollama and llama.cpp.
On first start-up, the buyer of an RTX Spark laptop picks an agent such as GitHub Copilot, Hermes, OpenClaw or Perplexity and installs it without a command line, Nvidia said [6]. The laptops launched Friday [5], so the specifications below are Nvidia's own.
Start with memory. The model needs 24GB [2] and the largest configuration offers up to 128GB of shared memory [8], so the model takes 18.75% of the biggest machine [14]. Nvidia says it runs on supported RTX Spark configurations [16]. TNW's report relays those claims. It does not say which configurations count as supported or what memory size the $2,599 starting price buys [11].
At 74 billion parameters in 24GB, the model works out to about 2.6 bits per parameter, if the 24GB holds only weights [13].
Microsoft describes the same machines as hybrid. "We're building Windows to be the platform for hybrid intelligence, where AI can run locally when it makes sense and reach the cloud when needed," said Yusuf Mehdi, executive vice president at Microsoft [12]. The tooling is built for users who send part of the work out [9]. Nvidia's open source Switchyard router lets agents split work between local and cloud models to cut cloud use [9].
We'd run a small pilot before standardising on anything. The trade-off is money and time: each seat starts at $2,599 [11], and the evidence the pilot starts from is the vendor's.
The forcing function we'd use is a two-by-two. One axis is whether the work must stay on the device. The other is whether the 24GB model completes your most common agent tasks when run on a single unit.
- Work must stay, model completes the tasks: run the pilot locally. - Work must stay, model fails: wait, because the cloud is not an option and the machine is not ready. - Work may leave, model completes the tasks: run locally where that cuts cloud use, and route the rest. - Work may leave, model fails: keep it in the cloud.
What to watch
- Whether Nvidia or a laptop maker states which memory size the $2,599 price buys and which configurations support Nemotron 3.5 Bolt.
- Whether Hermes ships Switchyard routing and OpenShell controls by the end of October, as Nvidia said it would.
- Independent tests of Nemotron 3.5 Bolt on agent tasks once reviewers have RTX Spark units.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence35
- Adoption
- Not scored
- Hype gap+30
- Incentives70
- Confidence60
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Nvidia's Nemotron 3.5 Bolt has 74 billion parameters, with 9 billion active at a time.
- [2]
Nemotron 3.5 Bolt is squeezed into 24GB of memory.
- [3]
Nvidia released Nemotron 3.5 Bolt on Thursday.
- [4]
Nemotron 3.5 Bolt's weights, training data and recipes are open, and it works with llama.cpp, Ollama, LM Studio, Unsloth and vLLM.
- [5]
Nvidia's RTX Spark laptops launched on Friday and are built to run AI agents on the laptop itself.
- [6]
Users can install agents such as GitHub Copilot, Hermes, OpenClaw and Perplexity from the first start-up, without a command line, Nvidia said.
- [7]
Each RTX Spark laptop runs on a single superchip that pairs a Blackwell RTX graphics chip with a Grace processor, with up to 1 petaflop of AI performance.
- [8]
The RTX Spark laptops have up to 128GB of shared memory.
- [9]
Nvidia's open source Switchyard model router lets agents split work between local and cloud models, to cut cloud use.
- [10]
Nvidia's OpenShell runtime and Microsoft Execution Containers let users control what agents are allowed to do; OpenClaw and GitHub Copilot already support Microsoft's containers.
- [11]
The laptops come from six makers (ASUS, Dell, HP, Lenovo, Microsoft and MSI); prices start at $2,599 and vary by region, Nvidia said.
- [12]
"We're building Windows to be the platform for hybrid intelligence, where AI can run locally when it makes sense and reach the cloud when needed."
- [13]
74 billion parameters in 24GB is about 2.6 bits per parameter, assuming the 24GB holds only weights and GB means 10^9 bytes.
- [14]
The 24GB model takes 18.75% of the 128GB top configuration.
- [15]
A ten-seat pilot at the $2,599 starting price costs at least $25,990 in hardware.
- [16]
Nvidia says Nemotron 3.5 Bolt runs on supported RTX Spark configurations.
- [17]
Hermes will add Switchyard later in October and will add OpenShell controls by the end of October, Nvidia said.
Get this ledger inside ChatGPT. Ask about any of our stories and the answer brings its figure, each claim’s verdict and the sources. Free, no account.
Sources
2 independent publishers whose own reporting we read for this story.
- techrepublic.comNvidia Launches Three Nemotron AI Models, Including Two for Local AI Agents
1 article · October 16, 2026
- thenextweb.comNvidia’s RTX Spark laptops are built to run AI agents locally
1 article · October 16, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.