← all repositories
peonist-ai/halogen-flash-server

One GPU, one model, zero fallbacks: a 4× faster LLM server

It rewrites every kernel for AMD Strix Halo and a single 179B-parameter MoE family to beat general-purpose engines without spending fewer bits.

halogen-flash-server
Collecting fresh signals — velocity needs a few days of history.
collecting data…
star history

What it does halogen-flash-server is an OpenAI-compatible inference host for Qwen3.8-Flash-Next that runs exclusively on AMD Strix Halo (gfx1151). It discards general-purpose runtimes in favor of kernels hand-written for that specific GPU and model family, with the stated goal of shrinking wall-clock time on long prompts. Since 0.7.0 it can also ingest off-the-shelf llama.cpp GGUF files and push them through the same specialized kernels.

The interesting bit The engine treats portability as a liability: there are no abstraction layers, no fallback paths, and no support for other silicon. Speculative decoding is verified to be byte-identical to greedy decode at temperature 0 on every release, using both the model’s own draft head and the request’s prompt text as draft sources—so it functions as a pure speed optimization rather than a quality trade.

Key highlights

  • Claims roughly 4× faster end-to-end latency than published alternatives for this model on Strix Halo, while running at 5.53 bits per weight—higher precision than most competing runs.
  • Exposes an OpenAI-shaped API (/v1) with optional vision, explicit 400 errors for unsupported sampling fields, and server-side defaults that can be set without restarting.
  • Opens existing llama.cpp GGUFs directly—such as unsloth’s UD-IQ4_XS—with no conversion step.
  • Supports up to 1M token context, attention budgets, and a preview of composable context.
  • Can operate fully offline after weights are cached, and optionally splits the API frontend from the inference backend.

Caveats

  • The decode speed comparison is deliberately softer than the prefill claim; the README notes that two of three competitors do not clearly state whether speculative decoding was enabled at 32K context.
  • The API and engine containers must use the exact same image tag, or they may miscommunicate features such as image placeholders.
  • It is welded to one GPU architecture and one model family; if you do not own Strix Halo, there is no compatibility path.

Verdict If you have Strix Halo hardware and want the lowest possible local latency for this specific 179B-parameter MoE, the project is built precisely for you. Anyone without that exact AMD chip should ignore it—the authors are explicit that portability was the price of admission.

Frequently asked

What is peonist-ai/halogen-flash-server?
It rewrites every kernel for AMD Strix Halo and a single 179B-parameter MoE family to beat general-purpose engines without spending fewer bits.
Is halogen-flash-server open source?
Yes — peonist-ai/halogen-flash-server is an open-source project tracked on heatdrop.
What language is halogen-flash-server written in?
peonist-ai/halogen-flash-server is primarily written in Shell.
How popular is halogen-flash-server?
peonist-ai/halogen-flash-server has 530 stars on GitHub.
Where can I find halogen-flash-server?
peonist-ai/halogen-flash-server is on GitHub at https://github.com/peonist-ai/halogen-flash-server.

heatdrop uses Google Analytics to see which pages get read — nothing else. Your call. How we handle data.