halogen-flash-server

halogen™ is the fastest way to run Qwen3.8-Flash-Next on AMD Strix Halo (gfx1151), and it is built for agents. One GPU, one model family, every kernel written for this exact silicon, shipped as a container with an OpenAI-compatible API.

Using it

The server speaks the OpenAI API at /v1, and /health is the authoritative account of what the build you are running supports: the sampling fields, whether images are accepted, the token budget aliases and the current default, and the tool-call wire format. What follows is the part worth reading first.

The documentation

Start here if you are new: Quickstart, then Give it a machine of its own: what the server holds and what that leaves for anything else.

Driving it from a client:

• API reference: every endpoint, conventions, status codes
• Codex and the Responses API
• Claude Code and the Anthropic Messages API
• From an agent harness
• Sessions: keep, fork and join
• Sampling · Images · Token budgets · Using it as a classifier

Configuring it:

• Configuration: cache modes, the KV pool, the prompt cache on disk, 1M context, the attention budget, composable context, an API key, request records
• Precision: what you get, and how to trade it
• Bring your own GGUF
• Small models on the NPU and the NPU host details

Understanding the numbers:

• Measured: every published number, both checkpoints, with the conditions each was taken under
• Quality: what is measured, and what is not
• Measuring a checkpoint: verify, inspect, ppl, niah as modes of the image

When something is wrong:

• Troubleshooting
• What this release is not

Reference:

• docs/FLAGS.md: every HALOGEN_* environment flag
• docs/QUANT.md: the shipped checkpoint's precision, by tensor family
• docs/CLASSIFY.md: the classify endpoint
• CHANGELOG: every release, what changed, and who reported it

For AI agents: AGENTS.md is the short playbook for deploying, driving or debugging the server, and llms.txt maps every documentation page in one fetch.