GTC Berlin: 8-unit DGX Spark clusters, Windows DGX Station

What was announced?

At GTC Berlin, NVIDIA showed its full local AI hardware lineup (NVIDIA). RTX Spark, launched last week, targets personal agents on Windows PCs. DGX Spark is for developers on Linux, and DGX Station serves enterprise teams running frontier workloads. We covered the RTX Spark laptops separately, so this piece collects the updates that matter to developers.

DGX Spark: smaller box, bigger clusters

A 64GB DGX Spark starts rolling out from partners on Friday, October 23. NVIDIA says it runs Nemotron 3.5 Bolt 74B, Qwen3.8 27B and Gemma 4 26B-A4B. Clusters can now include up to eight units in manual configuration, giving 1TB of unified memory, enough for frontier-class models like GLM 5.3. NVIDIA pitches this for small teams running concurrent sessions against one shared model.

NVIDIA Sync Model Launcher arrives later this month. In a few clicks it downloads a model and container suited to the configuration and connects the model to OpenCode. The October software update adds a Wi-Fi Assistant, while DGX Spark Skills bring troubleshooting and maintenance guidance through OpenCode.

DGX Station: Windows and a team inference service

DGX Station offers 748GB of coherent memory and up to 20 petaflops of AI performance. NVIDIA says it can serve more than 30 users on GLM 5.3 at speeds comparable to cloud models. Seven new IT management integrations include Microsoft Intune, Splunk and Tanium.

DGX Station for Windows is expected from system makers in December. It supports engineering software from Siemens, Dassault Systèmes, MathWorks, Synopsys and others, plus agents from Hermes Agent, Perplexity and Aible. NVIDIA Cloud Functions is coming in early access to Linux-based DGX Station. Teams register local model endpoints in a shared catalog and reach them through one OpenAI-compatible entry point. The model behind an endpoint can be swapped without changing application code. NVIDIA also released three open source robotics reference workflows.

Recursive Flint and new models

Recursive Superintelligence released Recursive Flint, a local coding model. It is a 40 billion-parameter mixture-of-experts model with 2.5 billion active parameters. NVIDIA says it generates over 220 tokens per second on a 64GB RTX Spark. The company claims best-in-class scores on SWE-Bench Pro v2, Terminal-Bench 2.1 and SWE-bench Verified without publishing the numbers in the post. Open weights are on Hugging Face under Apache 2.0.

Day-zero support covers Nemotron Bolt E4B, which runs in 4GB of memory and ships October 23, plus Liquid AI's LFM2.5-24B-A2B, MiniMax M3.1 Flash and the FLUX.3 Klein image model, which runs on GPUs with 8GB of VRAM. NVIDIA also released a collection of FP4-quantized checkpoints for local hardware, starting with Qwen3.8-Flash-Next-FP4-Local.

Faster llama.cpp and vLLM

llama.cpp gains MTP speculative decoding for Qwen 3.8 Flash Next, making agentic tasks up to 1.4x faster on an RTX PRO 6000 Blackwell. In vLLM, Qwen3.8 27B runs up to 1.3x faster on DGX Spark, and GLM 5.3 Flash gets up to 1.2x higher throughput on DGX Station. These are NVIDIA's own figures.

The takeaway for developers: local inference is becoming a team service. A Spark cluster or a DGX Station behind an OpenAI-compatible endpoint can replace part of a cloud API bill where data has to stay in house.