Happy New Year! Looking forward to a new year of building and learning on DGX Spark!
We’re kicking off 2026 with fresh updates across DGX Spark, including new performance improvements announced at CES and several new playbooks now available.
⭐ CES Highlights: Major performance gains announced during the CES keynote
Some of these optimizations are already integrated into existing playbooks—try them out on your Spark today.
⚙️ Updated Performance Improvements
-
TensorRT‑LLM: Performance gain in the latest container with improved two-Spark workflow
-
Speculative Decoding: New EAGLE-3 speculative decoding example running GPT-OSS-120B
If you visited CES, you may have seen additional demos on the show floor — we’re actively working on publishing these as new playbooks. Stay tuned for release announcements.
🆕 New Playbooks Available
-
Nemotron 3 Nano: Run NVIDIA’s efficient 30B-parameter MoE model locally for LLM experimentation.
-
Live VLM WebUI: Stream webcam input into vision-language models for real-time analysis, with GPU utilization.
-
Isaac Sim / Lab: Build and train robotics applications using GPU-accelerated simulation and reinforcement learning.
-
SGLang and vLLM inference playbooks: Now include a clear model support matrix showing tested and supported models and quantization options.
-
GPU-accelerated quantitative portfolio optimization and single cell RNA sequencing playbooks: Workflows with minimal code changes compared to CPU implementations.
-
Fine-tune with PyTorch: Distributed fine-tuning across two DGX Spark systems for LLMs up to 70B parameters using FSDP and LoRA.
Please continue sharing your feedback and/or latest projects in the forum or email spark-playbook-feedback@nvidia.com!