NVIDIA Expands RTX Spark Ecosystem With Nemotron 3.5 Transcribe, Liquid AI Models, and Agent Resource Controls

NVIDIA has announced a suite of local artificial intelligence software, open-weight models, and developer tools for its RTX Spark Windows laptop platform. The updates introduce speech, multi-modal, and agent-centric models—headlined by NVIDIA Nemotron 3.5 Transcribe and Liquid AI’s LFM3 family—alongside fine-grained GPU resource allocation controls inside the NVIDIA App.

RTX Spark systems, currently manufactured by ASUS, Dell, HP, Lenovo, Microsoft, and MSI starting at $2,599, combine an NVIDIA Blackwell RTX GPU and an NVIDIA Grace CPU on a single superchip. The systems deliver up to 128GB of unified memory, up to 1 petaflop of FP4 AI performance, and form factors measuring as thin as 14mm and weighing as light as three pounds.

Nemotron 3.5 Transcribe Brings Promptable Audio to Local Silicon

The focal point of NVIDIA’s software release is Nemotron 3.5 Transcribe, an open-weight speech model designed to combine transcription, speaker attribution, translation, and audio query responses into a single model. Rather than stringing together multiple specialized pipelines, developers can prompt the model using natural language commands, such as requesting a transcription with speaker labels, translating audio to English, or listing decisions made in a recorded meeting.

Key specifications and benchmark results for Nemotron 3.5 Transcribe include:

• Scale and Efficiency: Utilizes approximately 4.3 billion active parameters—80% fewer parameters than the current No. 1 open-weight model.
• Benchmarks: Ranks No. 1 on Hugging Face Open ASR and No. 2 among open-weight models on Artificial Analysis.
• Language Support: Handles 40 languages, including seven Indic languages, with customization options for domain-specific formatting and vocabulary across medical, finance, and customer service fields.
• Deployment Options: Available on Hugging Face now, with quantized NVFP4 versions scheduled for deployment via NeMo-Speech.cpp and vLLM across RTX Spark, DGX Spark, Jetson AGX Thor, and Blackwell data center GPUs.

Liquid AI LFM3 and Compact Agent Models

NVIDIA also rolled out day-zero support for Liquid AI’s LFM3 family of Liquid Foundation Models. Spanning text, vision, audio, omni, and decision tasks on a shared backbone, these open-weight models can be run and fine-tuned locally on RTX Spark using llama.cpp and vLLM. Liquid AI’s 1.7-billion-parameter base model notably outperforms its earlier 2.6-billion-parameter architecture, supporting autonomous agents in Hermes Agent, Pi, and OpenClaw to process screen inputs, manage email and calendars, and parse technical documents.

To broaden lightweight local agent deployments, NVIDIA also highlighted:

• Nemotron 3.5 Bolt E4B: Requires as little as 4GB of memory for multi-step agent actions and conversational tasks across tools like Ollama, Unsloth, LM Studio, and vLLM. A larger Nemotron 3.5 Bolt 74B-A9B (24GB) is tailored for heavier local workloads.
• FLUX 3 Image Klein: A 5.6-billion-parameter open-weights image generation and editing model from Black Forest Labs, deployable on RTX GPUs with 8GB of VRAM and customizable using NVIDIA NeMo AutoModel recipes.

System Resource Splitting and Simulation Workflows

To support systems running autonomous background agents while handling active user tasks, the NVIDIA App has introduced a resource management feature under System > Advanced. Users can shift GPU compute balance toward active foreground software or grant higher priority to background-running agents. The capability depends on Windows Hardware-Accelerated GPU Scheduling being active.

In robotics, NVIDIA outlined how RTX Spark was used during a GTC Berlin keynote demonstration to train Hugging Face's Microduck robot. In that workflow, Hermes Agent and Nemotron 3.5 Bolt planned an obstacle course, Astra generated 3D assets, NVIDIA Isaac Lab trained the navigation policy, and RTX Spark handled policy validation inside NVIDIA Isaac Sim.

For standalone workstation setups, NVIDIA confirmed that manufacturer partners are also shipping the desktop DGX Spark 64GB, equipped with the GB10 Grace Blackwell Superchip running DGX OS for edge fine-tuning and inference.