Spark-Studio

Overview

Spark-Studio is a friendly, LAN-ready dashboard designed for DGX Spark systems to manage local model serving across engines like vLLM, SGLang, llama.cpp, and WebGPU. It features automated setup wizards, hardware-aware recipe validation, live logs, benchmarking, and autonomous agent integration (Claude Code/Codex) to diagnose and patch failing or slow inference recipes automatically.

Capabilities

• ▸First-run setup wizard with hardware profiling
• ▸Multi-engine dashboard (vLLM, SGLang, llama.cpp, WebGPU)
• ▸Autonomous recipe debugging and auto-fix/retry loops
• ▸Speed optimization tuning loops with rollback capability
• ▸Hardware-aware recipe fit checking and compatibility analysis
• ▸Built-in chat, Monaco-based recipe editor, and markdown benchmarking reports
• ▸Tool evaluation benchmark suite for deterministic capability testing

Best for

Launching and monitoring local LLMs with one click on DGX Spark hardware, Diagnosing and automatically fixing broken inference recipes using AI agents, Benchmarking model speed (tokens/sec, TTFT) and evaluating tool-calling capabilities, Managing local Hugging Face model caches and translating model launch commands into YAML recipes

Works with

Security considerations

The dashboard binds to all network interfaces without authentication by default, allowing anyone on the local network full control including interactive shell access. It should only be run on trusted local networks.

Categories