# AI use cases for DGX Spark Qwen38 (hasso5703/dgx-spark-qwen38) Repo: [hasso5703/dgx-spark-qwen38](https://github.com/hasso5703/dgx-spark-qwen38) | Use cases: 4 | Page: https://dedupedai.com/use-cases/hasso5703-dgx-spark-qwen38/ A one-command installer that sets up a private AI server on a DGX Spark or ASUS Ascent GX10, serving local text, image, video and speech models through a cockpit and OpenAI- and Anthropic-compatible APIs. This file holds every use case of this repo. Each one starts with a "## " heading and ends with its URL. - Run a private coding agent on your own DGX Spark: https://dedupedai.com/use-cases/hasso5703-dgx-spark-qwen38/private-coding-agent-on-spark/ - Point Claude Code or Cursor at a local model: https://dedupedai.com/use-cases/hasso5703-dgx-spark-qwen38/claude-code-on-local-model/ - Make images and videos with sound on your own box: https://dedupedai.com/use-cases/hasso5703-dgx-spark-qwen38/generate-images-and-videos-offline/ - Read long confidential files in one 1M-token window: https://dedupedai.com/use-cases/hasso5703-dgx-spark-qwen38/analyze-long-documents-locally/ ## How to run a private coding agent on your own DGX Spark with DGX Spark Qwen38 Repo: [hasso5703/dgx-spark-qwen38](https://github.com/hasso5703/dgx-spark-qwen38) | Difficulty: Intermediate | Time: 2 hours | Basis: From the README Install one stack and point opencode at a 27B model that never leaves the box. **Who it is for:** A developer with a DGX Spark or GX10 who wants a coding assistant that keeps source code on local hardware. **The problem:** Hosted coding assistants send prompts and code to a third party, and setting up a local model stack by hand is slow. **The outcome:** A boot-persistent serving stack with the Qwen3.8-27B lane running, and opencode already wired to the local model. ### Steps 1. Run `curl -fsSL https://raw.githubusercontent.com/hasso5703/dgx-spark-qwen38/main/get.sh | bash` on the box, or clone the repo and run `./install.sh` if you prefer to read it first. 2. Wait for the install to finish; the lane that serves is proved before the run ends. 3. Open opencode; the installer writes a ready-to-use provider config for the local model (see docs/opencode.md). 4. Ask opencode to read a small project and make one change, then check the result by running that project's tests. ### Prompt to paste into your agent ```text Use the repo hasso5703/dgx-spark-qwen38 on a DGX Spark. Run its install.sh, then send one coding prompt to the local stock Qwen3.8-27B lane through opencode. Expected output: a code answer produced by the local model with no external API call. Verify that the install reports the serving lane as proved and that the answer returns. ``` ### Caveat The README measures these speeds on an ASUS Ascent GX10 and says to re-measure on your own box with ./bench-matrix.sh. Risk, remote script: The one-liner pipes a script from GitHub straight into bash, so clone the repo and read install.sh first if you do not trust the source. Dual-use notice: only use this on software, sites and systems you own or are authorized to test. Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/ URL: https://dedupedai.com/use-cases/hasso5703-dgx-spark-qwen38/private-coding-agent-on-spark/ ## How to point Claude Code or Cursor at a local model with DGX Spark Qwen38 Repo: [hasso5703/dgx-spark-qwen38](https://github.com/hasso5703/dgx-spark-qwen38) | Difficulty: Intermediate | Time: 1 hour | Basis: From the README Use the Anthropic-compatible endpoint on port 30001 so agent CLIs run on your own box. **Who it is for:** A developer who already uses Claude Code or Cursor and wants to keep their agent sessions on a local DGX Spark. **The problem:** Agent CLIs usually need a hosted API key and send project context off the machine. **The outcome:** Claude Code or Cursor sends its requests to the local Qwen3.8 lane through the Anthropic protocol on port 30001. ### Steps 1. Run the one-liner install on the DGX Spark: `curl -fsSL https://raw.githubusercontent.com/hasso5703/dgx-spark-qwen38/main/get.sh | bash`. 2. Open docs/clients.md in the repo and find the Claude Code and Cursor sections. 3. Set the client's base URL to the box's address on port 30001 and use the local model name from the cockpit. (inferred, check the README) 4. Send a short test prompt from the client and confirm the reply comes from the local lane. ### Prompt to paste into your agent ```text Use the repo hasso5703/dgx-spark-qwen38 to connect claude-code on a laptop to the local Qwen3.8 lane on . Expected output: a reply to a short test prompt returned through the Anthropic protocol on port 30001. Verify that the request is answered by the box and that no hosted API key is used. ``` ### Caveat The README lists Claude Code and Cursor under docs/clients.md, so exact client settings should be taken from that file rather than from this summary. Dual-use notice: only use this on software, sites and systems you own or are authorized to test. Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/ URL: https://dedupedai.com/use-cases/hasso5703-dgx-spark-qwen38/claude-code-on-local-model/ ## How to make images and videos with sound on your own box with DGX Spark Qwen38 Repo: [hasso5703/dgx-spark-qwen38](https://github.com/hasso5703/dgx-spark-qwen38) | Difficulty: Intermediate | Time: 3 hours | Basis: From the README Run Qwen-Image 2.1 and MiniMax-H3 locally from the cockpit, with one Load button. **Who it is for:** A content creator or designer who wants image and video generation without uploading media to a hosted service. **The problem:** Hosted image and video tools upload prompts and assets, and local setups for media models are hard to run alongside text models. **The outcome:** A 1024x1024 image from Qwen-Image 2.1 and a video with sound from MiniMax-H3, both made on the box. ### Steps 1. Run the one-liner install on the DGX Spark; add `--all-weights` to a clone's install to build and prove every lane up front. 2. Open the cockpit in a browser on a laptop or phone. 3. Press Load on the Qwen-Image 2.1 lane and generate a 1024x1024 image. 4. Press Load on the MiniMax-H3 lane and generate a short video; expect a long run, the README's example took 25 min 51 s. ### Prompt to paste into your agent ```text Use the repo hasso5703/dgx-spark-qwen38 on a DGX Spark to load the Qwen-Image 2.1 lane and generate one 1024x1024 image from the prompt . Expected output: an image file produced locally in about the time the README reports with Turbo. Verify that the image is written on the box and that the cockpit shows the lane loaded. ``` ### Caveat Lanes take turns in one 128 GB pool, so the cockpit must unload one lane before the other can load comfortably. Dual-use notice: only use this on software, sites and systems you own or are authorized to test. Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/ URL: https://dedupedai.com/use-cases/hasso5703-dgx-spark-qwen38/generate-images-and-videos-offline/ ## How to read long confidential files in one 1M-token window with DGX Spark Qwen38 Repo: [hasso5703/dgx-spark-qwen38](https://github.com/hasso5703/dgx-spark-qwen38) | Difficulty: Intermediate | Time: 2 hours | Basis: From the README Send a large document to the local 27B model with a 1M-token context. **Who it is for:** An analyst, lawyer or researcher with long confidential files who cannot send them to a hosted model. **The problem:** Long files are split or summarized in pieces by most hosted tools, and confidential material should not leave the office. **The outcome:** A single long document is answered from the local Qwen3.8-27B lane with a 1M-token context. ### Steps 1. Run the one-liner install on the DGX Spark: `curl -fsSL https://raw.githubusercontent.com/hasso5703/dgx-spark-qwen38/main/get.sh | bash`. 2. Load the stock target, which is the default, with 1M context by default. 3. Read docs/context-1m.md to check how the limit is fitted to your boot before sending very large files. 4. Send the document through the OpenAI-compatible API with a question, and check the answer against a few passages you know. (inferred, check the README) ### Prompt to paste into your agent ```text Use the repo hasso5703/dgx-spark-qwen38 to send to the local stock Qwen3.8-27B lane through its OpenAI-compatible API. Expected output: a summary and answers to three questions about the file, generated on the box. Verify that three answers match passages you have checked by hand. ``` ### Caveat The README says the 1M context is fitted to each boot, so very large inputs may need the limits in docs/context-1m.md checked first. Dual-use notice: only use this on software, sites and systems you own or are authorized to test. Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/ URL: https://dedupedai.com/use-cases/hasso5703-dgx-spark-qwen38/analyze-long-documents-locally/