Bringing the Spark to Europe: Local Intelligence Powers Next Wave of Agents
Source
blogs.nvidia.com
Author
Gerardo Delgado
Date
Why it matters
Lists dated local-inferenceRunning a trained model to get answers — the phase where AI is actually used, as opposed to trained.Full definition → changes: DGX Spark 64GB on Oct 23, eight-node NVIDIA Sync, a new local coding model, and llama.cpp and vLLM speedups. Matters if you run coding agents on local hardware.
Key takeaways · AI-distilled
DGX Spark now supports clusters of up to eight units in manual configuration over built-in ConnectX-7 networking, giving 1TB of unified memory that NVIDIA says can run frontier-class open models like GLM 5.3 for a small team sharing one backend.
NVIDIA reports speculative decodingA speed trick where a small model drafts several tokens ahead and the big model verifies them in one pass, often doubling generation speed.Full definition → with MTP in llama.cpp makes agentic tasks with Qwen 3.8 Flash Next up to 1.4x faster on RTX PRO 6000 Blackwell, and vLLM changes make Qwen3.8 27B tasks up to 1.3x faster on DGX Spark.
A new 64GB DGX Spark ships from partners on Oct. 23, and NVIDIA released quantizationShrinking a model by storing its numbers less precisely — like rounding — so it runs faster and fits on smaller hardware, at a small quality cost.Full definition → FP4 checkpoints meant to fit frontier-class open models with room left for KV cacheThe memory a model keeps about text it has already read, so generating each new token doesn't require reprocessing the whole conversation.Full definition →, starting with Qwen3.8-Flash-Next on a single 128GB DGX Spark.
Terms in this piece · Glossary
inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.
speculative decoding — A speed trick where a small model drafts several tokens ahead and the big model verifies them in one pass, often doubling generation speed.
quantization — Shrinking a model by storing its numbers less precisely — like rounding — so it runs faster and fits on smaller hardware, at a small quality cost.
KV cache — The memory a model keeps about text it has already read, so generating each new token doesn't require reprocessing the whole conversation.
