Skip to content
Neural Daily

ToolsNews

NVIDIA local AI: DGX Spark clusters and Recursive Flint

At GTC Berlin NVIDIA refreshed its local AI lineup: a 64GB DGX Spark ships Friday, up to eight Sparks can now be clustered, and Recursive Flint arrives as an open local coding model.

By: Mikola Máté Patrik · · 3 min read

NVIDIA local AI lineup: a DGX Station tower, an RTX Spark laptop and a DGX Spark mini PC on a circuit-patterned floor
Image: NVIDIA · Press kit · Official blog social image (og:image) · Source

What changed?

On October 21 at GTC Berlin, NVIDIA announced updates across its local AI hardware (NVIDIA blog). We already covered RTX Spark laptops going on sale; this round is about developer machines and the software stack.

DGX Spark

  • DGX Spark 64GB: rolls out from partners on Friday, October 23. NVIDIA says it runs models such as Nemotron 3.5 Bolt 74B, Qwen3.8 27B and Gemma 4 26B-A4B.
  • Clusters: up to eight DGX Spark units can now be linked over built-in ConnectX-7 networking in manual configuration. That is 1TB of unified memory, which NVIDIA says is enough for frontier-class models like GLM 5.3.
  • NVIDIA Sync Model Launcher: coming later this month. A few clicks download a model and container suited to the setup and connect it to OpenCode.
  • October software update: a Wi-Fi Assistant for appliance mode plus updated DGX Spark Skills for troubleshooting and maintenance.

DGX Station

  • 748GB of coherent memory and up to 20 petaflops of AI compute. NVIDIA says one system can serve more than 30 users.
  • NVIDIA Cloud Functions in early access on Linux DGX Station: local model endpoints go into a shared catalog behind one OpenAI-compatible entry point, so the model can be swapped without changing application code.
  • Seven new management integrations, including Microsoft Intune, Splunk and Tanium.
  • DGX Station for Windows arrives from system partners in December, with support for industrial apps from Siemens, Dassault Systèmes, MathWorks and others.
  • Three open source robotics reference workflows built on Isaac Lab, Cosmos and GR00T 1.7.

Recursive Flint: a local coding model

Recursive Superintelligence released Recursive Flint, a 40B-parameter MoE model with 2.5B active parameters per token. NVIDIA says it generates more than 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, but the announcement gives no figures. Weights are on Hugging Face under Apache 2.0, and the model can be fine-tuned with NeMo Automodel (SFT, LoRA) or trained further with GRPO in NeMo RL.

Day-zero model support

Day-zero support covers Nemotron Bolt E4B (runs in 4GB of memory, out October 23), the FLUX.3 Klein image model (5.6B parameters, from 8GB VRAM), MiniMax M3.1 Flash and Liquid AI's LFM2.5-24B-A2B. NVIDIA also released a collection of RAM-efficient quantized checkpoints, starting with Qwen3.8-Flash-Next-FP4-Local for a single 128GB DGX Spark. llama.cpp and vLLM get new accelerations too.

Why it matters

Local inference is turning into a team-level option: eight Sparks or one Station can act as a shared, OpenAI-compatible backend for coding agents, keeping sensitive code off the cloud. Flint's performance numbers are vendor claims for now, so independent benchmarks are worth waiting for.

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles