InfrastructureCompute, chips & cloud
NVIDIA launches RTX Spark Windows-on-Arm laptops and Nemotron 3.5 Bolt model for local coding agents
NVIDIA's RTX Spark laptops went on sale October 16. They use the GB10 chip with up to 128GB of unified memory and ship alongside a 2-bit 74B mixture-of-experts model, pitched as one Windows-on-Arm PC for coding, testing and deploying agents.
From the NVIDIA Generative AI article of October 16, 2026 and cited sources: RTX Spark puts the GB10 (N1X) chip from DGX Spark into Windows PCs. [1] [2] [4]
SOURCES[1] Code, Test, and Deploy on One Machine with NVIDIA RTX Spark developer.nvidia.com[2] The Register: Benchmarked: Nvidia's RTX Spark notebook theregister.com[4] Tom's Hardware: Nvidia enters the Windows PC market with… tomshardware.com
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- Original source
- NVIDIA Generative AI article, company source
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- October 16, 2026 article published
- Covered
- Oct 17 edition
- Evidence
- 10 sources across 8 sites
- 1Code, Test, and Deploy on One Machine with NVIDIA RTX Sparkdeveloper.nvidia.com
- 2The Register: Benchmarked: Nvidia's RTX Spark notebooktheregister.com
- 3ServeTheHome: NVIDIA introduces RTX Sparkservethehome.com
- 4Tom's Hardware: Nvidia enters the Windows PC market with…tomshardware.com
- 5NVIDIA Blog: All Systems Goblogs.nvidia.com
- 6Hugging Face: NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GBhuggingface.co
Reach: not measured yet
RTX Spark puts the GB10 (N1X) chip from DGX Spark into Windows PCs. It pairs a Grace CPU with a Blackwell GPU, up to 128GB of unified memory and about 300 GB/s of bandwidth. The software stack includes the first PyTorch-CUDA build for Windows on Arm, sandboxed agent options and integrations such as GitHub Copilot and JetBrains Junie Local. [1] [2] [4]
Nemotron 3.5 Bolt is a hybrid Mamba-2/Transformer mixture-of-experts model with 74B total and 9B active parameters. Its expert weights are quantized to about 2 bits so the model fits in 24GB. It is released under the OpenMDW-1.1 license, but NVIDIA's model page and its Hugging Face card disagree on context length, giving 128K and 256K. [6] [8]
NVIDIA's figures of 4.6x faster agents, 2.7x faster fine-tuning and 1.8x faster compilation against Intel's Panther Lake are vendor-reported and have not been reproduced. In separate testing, The Register measured about 68 tok/s on Qwen3.6-35B and 94 tok/s in an agentic test on an HP laptop. That beat AMD's Strix Halo and trailed DGX Spark slightly. [1] [2]
Measured throughput shows a laptop can run mid-size models locally at usable speeds. Whether that justifies $2,600 to $6,229 in hardware over cloud agent spending is still unproven.
Read the full assessment
Teams need to test their own workloads, Arm-native dependencies and performance under sustained load.
Executive brief
NVIDIA says its Windows-on-Arm RTX Spark PCs run coding agents 4.6x faster than Intel's Panther Lake. Nobody outside NVIDIA has reproduced that comparison. The first independent test does push against one of the claims: The Register found that the HP laptop it tested runs too hot for comfortable fine-tuning, yet NVIDIA claims a 2.7x fine-tuning advantage. The platform uses the same GB10 chip as DGX Spark, up to 128GB of unified memory, and a new 2-bit, 74B mixture-of-experts model (Nemotron 3.5 Bolt). The pitch is to code, test and deploy agents on one PC. Systems start around $2,600.
What changed and event timeline
RTX Spark unveiled
The chip, codenamed N1X, was announced before Computex. It pairs a 20-core Grace CPU with a Blackwell GPU and was slated to ship in the fall from ASUS, Dell, HP, Lenovo, Surface and MSI (;).
Software ecosystem announced
NVIDIA listed launch software: Nemotron Bolt models, NVIDIA PAIR leaving beta, and a claim that 97% of the top 15,000 Python wheels are ready ().
Nemotron 3.5 Bolt 74B released
The 74B-total, 9B-active MoE model went up on Hugging Face under the OpenMDW-1.1 license, sized to fit in 24GB ().
Laptops launch; reviews published
Laptops went on sale, and The Register and LTT Labs published reviews the same day (;).
Developer pitch published
NVIDIA's developer blog published its benchmarks against Intel, Microsoft toolchain updates, and agent integrations ().
Desktops due
Small-form-factor RTX Spark desktops, pitched for agents that run 24/7, are scheduled for November ().
Capabilities and access
- Chip: GB10 / N1X system-on-chip with up to 128GB unified memory and a claimed 1 petaflop of FP4 compute (NVIDIA Developer).
- Configurations: 18 CPU cores with 5,120 CUDA cores, or 20 cores with 6,144 CUDA cores. Memory bandwidth is about 300 GB/s (The Register).
- The first PyTorch-CUDA build for Windows on Arm. Python 3.15 Arm-native by default. More than 2,000 Conda packages.
Read the full section
- Chip: GB10 / N1X system-on-chip with up to 128GB unified memory and a claimed 1 petaflop of FP4 compute (NVIDIA Developer).
- Configurations: 18 CPU cores with 5,120 CUDA cores, or 20 cores with 6,144 CUDA cores. Memory bandwidth is about 300 GB/s (The Register).
- Software: Windows Developer Config or WSL; the first PyTorch-CUDA build for Windows on Arm; Python 3.15 Arm-native by default; more than 2,000 Conda packages.
- Agents: GitHub Copilot via Windows ML with a llama.cpp backend, Pi, JetBrains Junie Local, Hermes and OpenClaw. Sandboxing comes from OpenShell or Microsoft Execution Containers (NVIDIA Developer).
Technical analysis for researchers and developers
- Bolt 74B design: a hybrid Mamba-2/Transformer MoE with a C-RADIO v4-H vision encoder. Expert weights use IQ2 (about 2-bit) quantization with quantization-aware training.
- Deployment notes: the model card recommends running one sequence at a time on 24GB targets.
- Porting: apps can be built natively for Arm64 or migrated piece by piece with Arm64EC.
Read the full section
- Bolt 74B design: a hybrid Mamba-2/Transformer MoE with a C-RADIO v4-H vision encoder. Expert weights use IQ2 (about 2-bit) quantization with quantization-aware training; the Mamba layers stay in NVFP4 and the embeddings in Q8_0 (Hugging Face). NVIDIA's model page targets at least 80 tokens/s decoding (NVIDIA Research).
- Deployment notes: the model card recommends running one sequence at a time on 24GB targets. Prompts, images and tool output all share the context window.
- Porting: apps can be built natively for Arm64 or migrated piece by piece with Arm64EC.
- NVIDIA's benchmark setup: AA-AgentPerf-local with Qwen3.8-27B, Unsloth LoRA fine-tuning on Llama-3.1-8B, and a Linux kernel build via Phoronix. The baseline is an Intel Core Ultra X9 388H set to 1.0.
Claims and evidence
- 4.6x / 2.7x / 1.8x over Panther Lake (agents, fine-tuning, compilation): these are NVIDIA's own figures (NVIDIA Developer). None of the reviewed sources reproduces them.
- Independent throughput: on the HP OmniBook Ultra 16, The Register measured 67.64 tok/s with Qwen3.6-35B and a 208 ms time to first token.
- Bolt model scores: GPQA Diamond 86.43, Terminal-Bench 2.1 50.98, BrowseComp 66.75. These come from NVIDIA's model card (Hugging Face).
Read the full section
- 4.6x / 2.7x / 1.8x over Panther Lake (agents, fine-tuning, compilation): these are NVIDIA's own figures (NVIDIA Developer). None of the reviewed sources reproduces them.
- Independent throughput: on the HP OmniBook Ultra 16, The Register measured 67.64 tok/s with Qwen3.6-35B and a 208 ms time to first token. In the agentic test it measured 94.10 tok/s. It reports the laptop outperformed AMD's Strix Halo, with DGX Spark slightly ahead (The Register).
- Bolt model scores: GPQA Diamond 86.43, Terminal-Bench 2.1 50.98, BrowseComp 66.75. These come from NVIDIA's model card (Hugging Face).
- "Up to 200x faster" decision models: NVIDIA gives no benchmark or methodology for this figure.
Context and prior work
- RTX Spark reuses the GB10 chip from the Linux-based DGX Spark, which Tom's Hardware found beat AMD's Ryzen AI Max+ 395.
- It enters a Windows-on-Arm market that Qualcomm opened; LTT Labs' review headline frames it exactly that way (LTT Labs).
- Microsoft is described as a co-developer, adding Windows security features for on-device agents and kernel optimizations that were not available on Snapdragon (Tom's Hardware).
Read the full section
- RTX Spark reuses the GB10 chip from the Linux-based DGX Spark, which Tom's Hardware found beat AMD's Ryzen AI Max+ 395.
- It enters a Windows-on-Arm market that Qualcomm opened; LTT Labs' review headline frames it exactly that way (LTT Labs).
- Microsoft is described as a co-developer, adding Windows security features for on-device agents and kernel optimizations that were not available on Snapdragon (Tom's Hardware).
Limitations, safety and contested findings
- Thermals: The Register saw the GPU pass 90°C under sustained load, with graphical artifacts.
- Gaming: the whole chip shares limited memory bandwidth, which hurts graphics workloads (Tom's Hardware).
- Context length conflict: NVIDIA's model page cites a 128K window (NVIDIA Research), while the Hugging Face card describes 256K.
Read the full section
- Thermals: The Register saw the GPU pass 90°C under sustained load, with graphical artifacts. It judged the laptop short on thermal headroom for LLM fine-tuning, which cuts against NVIDIA's fine-tuning claim (The Register).
- Gaming: the whole chip shares limited memory bandwidth, which hurts graphics workloads (Tom's Hardware).
- Context length conflict: NVIDIA's model page cites a 128K window (NVIDIA Research), while the Hugging Face card describes 256K.
- Data openness: "fully open" training data excludes the exact post-training data mix, which the model card does not disclose. The card also says post-training used synthetic data from DeepSeek, Qwen and GPT-OSS models.
Business and practitioner implications
- Cost: the tested HP configuration costs $6,229, and entry systems are about $2,600 (The Register).
- Routing: local models plus hybrid routers such as HydraFusion could reduce cloud API use. NVIDIA has not published savings figures.
- Porting: Windows-on-Arm apps need a dependency audit first. Teams should test native code paths, not just Python wheels.
Read the full section
- Cost: the tested HP configuration costs $6,229, and entry systems are about $2,600 (The Register). Teams should weigh that against what they currently spend on cloud agent tokens.
- Routing: local models plus hybrid routers such as HydraFusion could reduce cloud API use. NVIDIA has not published savings figures.
- Porting: Windows-on-Arm apps need a dependency audit first. Teams should test native code paths, not just Python wheels.
- Hardware choice: for sustained fine-tuning, the November desktops or DGX Spark may suit better than thin laptops, though no source has tested the desktops yet.
The source trail.
Sources (10)
Code, Test, and Deploy on One Machine with NVIDIA RTX Spark
Article text retrieved; extracted text may omit tables or interactive elements.
developer.nvidia.com