Deepseek 4.1 Flash on 2 DGX Sparks changed my life

Summary

Alex Finn buys a second Nvidia DGX Spark and connects it to his existing unit to test how well a two-device local AI setup can run DeepSeek 4.1 Flash. Each Spark is described as having 128 GB of memory, for a total of 256 GB across the pair. The model download is around 400 GB, and the video presents the two Sparks as a way to run a larger local model than one device could handle alone.

Setup and Features

• Headless setup: Finn says the Spark can be configured without a monitor. It creates a Wi-Fi network for initial setup, after which he manages it from a Mac Studio.
• Connecting the devices: He links the Sparks with a cable so they can work together, then uses Tailscale to put the Sparks and his other computers on a private network.
• AI-agent-assisted configuration: He asks an agent he calls “Grokbot” to configure the second Spark and download the model. He also connects the local model to Hermes, creating a profile that lets him chat with it from his Mac Studio.
• Suggested workflow: For people unsure what model their hardware can run, Finn recommends asking an AI agent to inspect the computer and suggest a suitable model. The video also mentions Nvidia’s included recipe book as a resource for setup and local AI tasks, such as running models and training LoRAs.

Tests and Observations

• Basic chat: A short “Are you here?” test takes about eight seconds, according to Finn.
• Coding test: The model writes a single-file Snake game in under a minute. Finn says it took about a minute and a half to create and try it.
• More demanding task: He asks for a multi-level, top-down 3D roguelike inspired by Diablo, made with Three.js. He reports that it takes roughly 20–25 minutes. The model produces a playable game with enemies, upgrades between levels, and combat abilities. Finn also notes that it tested its own output and supplied screenshots.

These are demonstrations rather than controlled benchmarks. Finn is impressed by the results but acknowledges that the model is not at the cutting edge. He says he would still use higher-end cloud models for some complex programming work.

Main Takeaway

Finn argues that local models are useful not only as substitutes for paid cloud services, but also for frequent or continuous tasks where cloud usage limits can be restrictive. He calls this “ambient intelligence,” giving examples such as repeatedly reviewing code or searching for business opportunities.

Local inference avoids per-request cloud charges, though the hardware and electricity still have costs. Finn says a two-Spark setup is not necessary to get started: a Mac mini or another suitable computer can also run smaller local models.

Guide and Tutorial Points

1. Set up a Spark over its Wi-Fi connection, without attaching a monitor.
2. Connect multiple machines with Tailscale so they can reach one another.
3. Ask an AI agent to inspect the hardware, recommend a compatible model, and help install it.
4. Expose the local model through an API and connect it to a chat or agent interface such as Hermes.
5. Try local models for repeated, ongoing tasks, even if cloud models remain preferable for some demanding work.

Main Speaker and Sources

Alex Finn narrates the setup, tests, and conclusions. He also references a performance post by Ray Fernandez. The model’s responses and generated code are shown during the demonstrations. Micro Center is identified as the video sponsor.