Why This Site?
The NVIDIA DGX Spark ($7,999) has roughly 50,000 owners. Almost none have found good information about what to build with it. This site fills that gap with real, verified data — not speculation. Every guide and benchmark comes from a live, running GB10 Blackwell superchip.
What you'll find here:
- ✅ Step-by-step setup guides for models, frameworks, and tools
- ✅ Real performance benchmarks against cloud GPUs and consumer hardware
- ✅ Project ideas ranked by difficulty and impact
- ✅ Cost analyses that actually account for everything
- ✅ Honest limitations — what doesn't work, and why
Getting Started
What to Build First With Your DGX Spark: 10 Starter Projects
Just got your DGX Spark? Here's the prioritized list of projects that will teach you the most and give you the most value. From quick wins to ambitious builds.
Beginner GuideThe DGX Spark Homelab Setup Guide
Network configuration, power considerations, cooling, and everything you need to integrate the Spark into your homelab without blowing your budget.
Setup InfrastructureDGX Spark Maintenance: What to Watch Out For
Software updates, driver issues, thermal monitoring, and the known pain points that NVIDIA documentation conveniently omits.
Maintenance TroubleshootingBenchmarks
Running Qwen3.6:35B on DGX Spark — Full Setup & Benchmarks
Real performance data from a live GB10. Token throughput, memory utilization, first-token latency. Current results: 54.6 tok/s, 32GB VRAM.
Benchmark Qwen3.6 Live DataRunning vLLM on DGX Spark: Performance, Limits, Tips
How to squeeze every drop of performance from vLLM on the GB10 Blackwell. Architecture quirks, optimization settings, ARM64-specific gotchas.
Benchmark vLLM OptimizationBest Models to Run on DGX Spark (Ranked by Performance)
From 7B to 70B+ parameters — which models fit in 128GB unified memory, which run fast, and which choke. Ranked by speed, quality, and utility.
Benchmark Model RankingsProjects
How to Turn Your DGX Spark Into a Private AI API
Full walkthrough: set up Ollama + OpenRouter-compatible endpoints, configure HTTPS, expose your models. Your own inference infrastructure for pennies.
Project API ProductionBuilding a Crypto Trading Bot on the DGX Spark
Use your Spark's massive unified memory and Blackwell GPU for ML-enhanced trading. Full architecture: Freqtrade base, LightGBM signals, real-time inference.
Project Crypto MLDGX Spark as a Local RAG Server — Complete Setup
Build a private, offline RAG system for documents, research, and knowledge bases. Embed models, vector databases, and query pipelines — all locally.
Project RAG KnowledgeFine-Tuning Llama 3 on DGX Spark: What Fits, What Doesn't
Which models can be fine-tuned on 128GB unified memory, how to split CPU/GPU, quantization strategies, and the hard limits you'll hit.
Project Fine-tuning Llama 3DGX Spark for Computer Vision: What You Can Actually Do
Running vision models on the GB10: CLIP, SAM, video analysis, and the edge cases where the unified memory architecture actually shines.
Project Computer VisionTop 10 AI Projects People Build With the DGX Spark
A curated list of the most popular and most valuable DGX Spark projects from the community, with setup difficulty, cost, and time-to-value.
Roundup CommunityComparisons
DGX Spark vs RTX 4090 for LLM Inference
The $7999 vs $1600 showdown. Same Blackwell architecture, different form factors. Latency, throughput, memory limits, and the math on value.
Comparison RTX 4090 ValueDGX Spark vs Cloud GPU for AI Workloads — Cost Analysis
Real TCO breakdown: Spark vs AWS g5 vs GCP A100 vs Azure H100. Break-even at what token volume? Hidden costs nobody mentions.
Comparison Cloud TCODGX Spark vs Mac Studio M4 Max for AI
Unified memory battle: 128GB Blackwell vs 128GB Apple Silicon. Model sizes, inference speed, software ecosystem, and which makes more sense.
Comparison Mac Studio Unified MemoryLive Benchmark
Our DGX Spark is running live right now with Qwen3.6:35b-a3b (35B parameters, Q4_K_M quantized). Here are the latest verified metrics:
- Throughput: 54.6 tokens/sec
- Model VRAM: 32 GB (of 128 GB total)
- Architecture: GB10 Grace Blackwell Superchip
- Quantization: Q4_K_M (GGUF)