# AI use cases for Halogen Flash Server (peonist-ai/halogen-flash-server) Repo: [peonist-ai/halogen-flash-server](https://github.com/peonist-ai/halogen-flash-server) | Use cases: 5 | Page: https://dedupedai.com/use-cases/peonist-ai-halogen-flash-server/ halogen-flash-server is a Shell-based inference server that runs the Qwen3.8-Flash-Next model on AMD Strix Halo (gfx1151) GPUs through an OpenAI-compatible API, packaged as a container image. This file holds every use case of this repo. Each one starts with a "## " heading and ends with its URL. - Run a local Codex agent on a Strix Halo box: https://dedupedai.com/use-cases/peonist-ai-halogen-flash-server/run-local-codex-agent-on-strix-halo/ - Point Claude Code at a local halogen server: https://dedupedai.com/use-cases/peonist-ai-halogen-flash-server/run-local-claude-code-on-strix-halo/ - Resume a long agent session after a server restart: https://dedupedai.com/use-cases/peonist-ai-halogen-flash-server/keep-a-long-agent-session-warm-across-restarts/ - Benchmark decode and prefill on your own box: https://dedupedai.com/use-cases/peonist-ai-halogen-flash-server/measure-throughput-on-your-own-strix-halo/ - Use the server as a text classifier endpoint: https://dedupedai.com/use-cases/peonist-ai-halogen-flash-server/serve-a-local-text-classifier-endpoint/ ## How to run a local Codex agent on a Strix Halo box with Halogen Flash Server Repo: [peonist-ai/halogen-flash-server](https://github.com/peonist-ai/halogen-flash-server) | Difficulty: Intermediate | Time: 1 hour | Basis: From the README Serve Qwen3.8-Flash-Next locally and point Codex at it through the Responses API. **Who it is for:** A developer with a Strix Halo machine who wants a coding agent that runs on their own hardware instead of a hosted API. **The problem:** Coding agents usually send code and prompts to a hosted service, and the developer wants a local model that a standard agent client can drive. **The outcome:** Codex sends its requests to the halogen server on the local machine and gets answers from Qwen3.8-Flash-Next. ### Steps 1. Set the BIOS carve-out (UMA frame buffer) to its minimum, not Auto, as described in docs/host-setup.md. 2. Run `mkdir -p ~/halogen-models` 3. Run `podman run --rm -p 8731:8731 --device /dev/kfd --device /dev/dri --group-add keep-groups --ipc=host --ulimit memlock=-1:-1 -e HALOGEN_DOWNLOAD=peonist-ai/halogen-qwen3.8-flash-next -v ~/halogen-models:/models ghcr.io/peonist-ai/halogen-flash-server:0.18.12` (inferred, check the README) 4. Follow docs/codex.md to point Codex at the Responses API on the server. 5. Check the endpoint conventions and status codes in docs/api.md before sending a test request. ### Prompt to paste into your agent ```text Using the repo peonist-ai/halogen-flash-server, configure Codex on to use the local halogen server on port 8731, following docs/codex.md. Expected output: Codex completes a short test task against the local model and returns an answer. Verify that the server responds on port 8731 before running the test. ``` ### Caveat The image only runs on AMD Strix Halo (gfx1151) hardware, and the README says it has no fallback path for other GPUs. 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/peonist-ai-halogen-flash-server/run-local-codex-agent-on-strix-halo/ ## How to point Claude Code at a local halogen server with Halogen Flash Server Repo: [peonist-ai/halogen-flash-server](https://github.com/peonist-ai/halogen-flash-server) | Difficulty: Intermediate | Time: 1 hour | Basis: From the README Serve the model on your own box and route Claude Code through the Messages API. **Who it is for:** A developer who already uses Claude Code and wants its requests answered by a model running on their own Strix Halo machine. **The problem:** Claude Code needs a Messages API endpoint, and the developer wants that endpoint served from local hardware. **The outcome:** Claude Code sends its Messages API calls to the halogen server and receives answers from the local model. ### Steps 1. Set the BIOS carve-out (UMA frame buffer) to its minimum, as described in docs/host-setup.md. 2. Run `mkdir -p ~/halogen-models` 3. Run `podman run --rm -p 8731:8731 --device /dev/kfd --device /dev/dri --group-add keep-groups --ipc=host --ulimit memlock=-1:-1 -e HALOGEN_DOWNLOAD=peonist-ai/halogen-qwen3.8-flash-next -v ~/halogen-models:/models ghcr.io/peonist-ai/halogen-flash-server:0.18.12` (inferred, check the README) 4. Follow docs/claude-code.md to direct Claude Code to the Messages API on the server. 5. Send one small request as a test and confirm the reply, using docs/api.md for the expected responses. ### Prompt to paste into your agent ```text Using the repo peonist-ai/halogen-flash-server, configure Claude Code on to use the local halogen server on port 8731, following docs/claude-code.md. Expected output: a short Claude Code prompt returns a reply from the local model. Verify the server is listening on port 8731 before you test. ``` ### Caveat The README describes Claude Code support through the Messages API, but the exact client setup lives in docs/claude-code.md, which this summary does not reproduce. 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/peonist-ai-halogen-flash-server/run-local-claude-code-on-strix-halo/ ## How to resume a long agent session after a server restart with Halogen Flash Server Repo: [peonist-ai/halogen-flash-server](https://github.com/peonist-ai/halogen-flash-server) | Difficulty: Advanced | Time: 1 hour | Basis: From the README Use the on-disk prompt cache so a deep agent session picks up without a cold start. **Who it is for:** A developer running a long coding-agent session with tens of thousands of tokens of context who needs to restart the server without losing that progress. **The problem:** A cold first turn on a long prompt is slow, and a restart normally throws away the conversation state the agent built up. **The outcome:** The conversation resumes from the disk prompt cache after a restart instead of reprocessing the whole prompt. ### Steps 1. Read docs/configuration.md#prompt-cache-on-disk-resume-a-conversation-after-a-restart to learn how the disk cache is enabled. 2. Read docs/sessions.md to learn how to keep, fork, or join a session. 3. Start the server with the disk cache option set as described in docs/configuration.md, using the run command from the Quickstart. 4. Send a long multi-turn session through the API and let the cache fill. 5. Restart the server and send the next turn of the same session, then compare the first-token time with the cold first turn. ### Prompt to paste into your agent ```text Using the repo peonist-ai/halogen-flash-server, enable the on-disk prompt cache as described in docs/configuration.md and resume after a server restart. Expected output: the resumed session answers its next turn without reprocessing the full prompt. Verify by checking the time to first token against the first cold turn. ``` ### Caveat The README gives the cold first turn for a 34,000-token prompt as 21 s, so resume timing will depend on your own hardware and cache state. 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/peonist-ai-halogen-flash-server/keep-a-long-agent-session-warm-across-restarts/ ## How to benchmark decode and prefill on your own box with Halogen Flash Server Repo: [peonist-ai/halogen-flash-server](https://github.com/peonist-ai/halogen-flash-server) | Difficulty: Advanced | Time: 1 hour | Basis: From the README Reproduce the published throughput numbers on your own Strix Halo machine. **Who it is for:** A researcher or hardware tester who wants to check published speed claims against their own machine before trusting them. **The problem:** Published throughput figures are only useful if they can be reproduced on other hardware, and the team wants comparable numbers. **The outcome:** A set of prefill and decode measurements from the user's own machine, comparable with the figures in docs/performance.md. ### Steps 1. Set up the server on the Strix Halo box using the Quickstart in docs/getting-started.md. 2. Open docs/performance.md#served-throughput-end-to-end-over-http and follow the single command it gives for the prefill rows. 3. Run the decode measurement the same page describes, using the kind of text each row names. 4. Record the results with the host settings noted in docs/troubleshooting.md#the-host-settings-these-numbers-were-measured-on. 5. Post the command and the curve to the Discord community linked in the README if you want to compare with others. ### Prompt to paste into your agent ```text Using the repo peonist-ai/halogen-flash-server, run the prefill and decode benchmarks described in docs/performance.md on . Expected output: a table of prefill tok/s and decode tok/s for each prompt length and text type. Verify the host settings match docs/troubleshooting.md before comparing against the published 0.18.0 figures. ``` ### Caveat The README notes that the best decode case uses short repetitive prompts, so the peak tok/s figure should not be read as typical output speed. 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/peonist-ai-halogen-flash-server/measure-throughput-on-your-own-strix-halo/ ## How to use the server as a text classifier endpoint with Halogen Flash Server Repo: [peonist-ai/halogen-flash-server](https://github.com/peonist-ai/halogen-flash-server) | Difficulty: Intermediate | Time: 1 hour | Basis: From the README Send labeling and classification requests to the local model through the classify endpoint. **Who it is for:** An operations or data analyst who needs to sort incoming text into categories without sending it to an outside service. **The problem:** Sorting large batches of text by hand is slow, and the text may be too sensitive to send to a hosted classifier. **The outcome:** Each input text comes back with a category label produced by the local model. ### Steps 1. Start the server with the Quickstart run command, keeping the model on the local machine. 2. Read docs/classifier.md to learn the request shape for classification. 3. Read docs/CLASSIFY.md for the classify endpoint reference. 4. Send a small batch of sample texts and check the labels against a few examples you already know. 5. Scale the batch once the labels match your expectations. ### Prompt to paste into your agent ```text Using the repo peonist-ai/halogen-flash-server, send the texts in to the classify endpoint as described in docs/CLASSIFY.md. Expected output: one label per input text. Verify the labels on ten known examples before running the full batch. ``` ### Caveat The README points to docs/CLASSIFY.md for the endpoint, and this summary does not restate its request format, so check that page before building on it. 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/peonist-ai-halogen-flash-server/serve-a-local-text-classifier-endpoint/