NVIDIA launches RTX Spark for personal agents and DGX Spark 64GB on Oct. 23
NVIDIA used its GTC Berlin event to unveil a full line of local AI hardware. The RTX Spark laptop, released the week before the show, is marketed for personal agents on Windows PCs and offers up to 128GB of unified memory, 1 petaflop of AI compute and 1440p gaming at 100 fps. Underdog, a desktop‑agent app that runs entirely on‑device, is now available for download on RTX Spark.
Key points
- RTX Spark launches with up to 128GB memory, 1 petaflop AI compute and supports the Underdog desktop‑agent app
- DGX Spark 64GB becomes available on Oct. 23 and can cluster up to eight units for 1 TB unified memory
- Recursive Flint, a 40‑billion‑parameter coding LLM, is released with open weights and runs over 220 tokens/second on RTX Spark
The company also announced the DGX Spark 64GB system, which ships on Friday, Oct. 23, and can be linked with up to eight units for a 1 TB unified‑memory cluster. Developers can run frontier models such as Nemotron 3.5 Bolt 74B, Qwen3.8 27B and Gemma 4 26B‑A4B on the hardware. NVIDIA Sync will add a Model Launcher later this month to simplify model deployment. DGX Station, aimed at larger teams, provides 748GB of coherent memory, up to 20 petaflops of AI performance and will arrive for Windows in December.
Recursive Flint, a new 40‑billion‑parameter local coding LLM built by Recursive Superintelligence, is released with open weights under an Apache 2.0 license. It runs at over 220 tokens per second on a 64GB RTX Spark and scores highly on SWE‑Bench‑Pro v2 and related coding benchmarks. Day‑zero support was added for several other models, including Nemotron Bolt E4B 4GB and FLUX.3 Image Klein.
Model pages: GLM 5.3 → · FLUX 3 Image Klein →
The story so far
4 episodes →- NVIDIA launches RTX Spark for personal agents and DGX Spark 64GB on Oct. 23this story
Bringing the Spark to Europe: Local Intelligence Powers Next Wave of Agents
NVIDIA Blog · 21 October 2026
Loading the full article…
This text was published by NVIDIA Blog and written by Gerardo Delgado. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Hardware & Compute
All →- Intel and AMD collaborate on oneDNN optimizations for PyTorch · 2 src
- Nvidia says 1GW AI factory costs $50 billion, unveils DSX POD · 4 src
- Google Cloud and NVIDIA make A5X instances generally available next month · 2 src
- Nokia and Nvidia target 50% spectrum efficiency by 2027 · 3 src
- Nvidia expands Confidential Computing with Confidential Bare Metal for model builders · 1 src
Comments
via GitHub Discussions