In a video published October 31, Alex Ziskind reported that LM Studio generated 61 tokens per second on native Linux while an MSI Prestige N16 Flip AI Plus was unplugged, and 15 tokens per second when it was plugged in. He attributed the drop to missing support for NVIDIA Platform Controllers and Framework in Linux. The test concerns one RTX Spark N1X laptop and one workload. Watch Ziskind’s English-language test.
One MSI laptop’s Linux results changed when plugged in
Ziskind’s results put a practical limit on what this test says about Linux on RTX Spark: native Linux ran the LM Studio workload at 61 tokens per second unplugged and 15 when plugged in. He linked the slowdown to the absence of NVIDIA Platform Controllers and Framework support in Linux. Those figures describe his MSI Prestige N16 Flip AI Plus setup, not every RTX Spark laptop or Linux installation.
For Linux tools, Ziskind described Windows Subsystem for Linux (WSL), which runs a Linux environment within Windows, as the practical option in his setup. He reported 65.1 tokens per second on WSL in a separate GPTQ Qwen/OSS 120B test, close to the 65.4 tokens per second he reported on native Windows. He reported 61 tokens per second for that workload on DGX Spark. These are results from that particular test, not a general performance ranking.
NVIDIA’s October 28 AI updates
NVIDIA’s October 28 announcement covered Nemotron 3.5 Transcribe, Liquid AI’s LFM3 models and a new NVIDIA App control for balancing GPU resources between a foreground app and a background agent. NeoTeo’s earlier coverage of RTX Spark and DLSS 5 provides background on the platform’s gaming updates.
Nemotron 3.5 Transcribe is an open-weight speech model for transcription, speaker attribution, translation and questions about audio. NVIDIA says it covers 40 languages, including seven Indic languages. On October 28, NVIDIA said its weights were available on Hugging Face in BF16, FP8 and NVFP4 formats, while optimized quantized versions for local execution on RTX Spark and other named NVIDIA systems were coming soon. NVIDIA named NeMo-Speech.cpp and vLLM as deployment options for NVFP4 versions. NVIDIA’s announcement distinguishes the downloadable weights from those optimized local versions.
Liquid AI’s LFM3 family spans text, vision, audio, omni and decision models. NVIDIA said developers could run LFM3 locally on RTX Spark through llama.cpp and vLLM and fine-tune it locally. NVIDIA also said the new 1.7-billion-parameter base model outperformed its previous 2.6-billion-parameter base model.
NVIDIA App’s GPU-priority control
NVIDIA placed the resource-sharing control in NVIDIA App under System > Advanced. It lets users favor a foreground app or give more GPU resources to a background agent. NVIDIA says hardware-accelerated GPU scheduling must be enabled in Windows Settings for the control to work.
