GPU
Qwen3.8-Flash-Next
- Model id
halogen-qwen3.8-flash-next- Endpoint
/v1/chat/completions, /v1/completions, /v1/responses, /v1/messages- For
- The main model: chat, code, agents, long contexts.
The roster
Models
A request picks a model by its model. Every request that names no NPU model goes to the Flash model. Name the NPU models you want in HALOGEN_NPU_MODELS.
On the GPU
Every request that names no NPU model goes here.
On the NPU
NPU
Decide
decider-0.8b
- Model id
decider-0.8b- Endpoint
/v1/systemone (decisions)- For
- Give a text and a question with 2 to 10 options, get a probability for every option from one pass.
NPU
Decide
d3-lite, d3-nano
- Model id
d3-lite, d3-nano- Endpoint
/v1/systemone- For
- Decisions over the whole prompt, pictures included: up to 255 options, a yes or no, or a score.
NPU
Embed
qwen3-embedding-0.6b
- Model id
qwen3-embedding-0.6b- Endpoint
/v1/embeddings- For
- Embeddings, 1024 floats each, unit length.
NPU
Embed
embeddinggemma-2
- Model id
embeddinggemma-2- Endpoint
/v1/embeddings- For
- Text, pictures and sound in one space (768 floats each).
NPU
Rerank
qwen3-reranker-0.6b
- Model id
qwen3-reranker-0.6b- Endpoint
/v1/rerank- For
- Score documents against a query.
NPU
Moderate
qwen3guard-gen-0.6b
- Model id
qwen3guard-gen-0.6b- Endpoint
/v1/moderations- For
- Is a prompt, or a reply in its conversation, safe?
NPU
Moderate
qwen3guard-stream-0.6b
- Model id
qwen3guard-stream-0.6b- Endpoint
a streaming guard on Flash replies- For
- Labels a reply from the Flash model as it is written.
NPU
Text
qwen3.5-2b
- Model id
qwen3.5-2b- Endpoint
/v1/chat/completions- For
- Summaries and other short jobs, thinking off.
NPU
Images
flux2-klein-4b
- Model id
flux2-klein-4b- Endpoint
/v1/images/generations, /v1/images/edits- For
- Icons, diagrams, illustrations and placeholders from 256x256 up to 1024x1024; edits of up to four pictures.
NPU
Speech to text
qwen3-asr-0.6b, qwen3-asr-1.7b
- Model id
qwen3-asr-0.6b, qwen3-asr-1.7b- Endpoint
/v1/audio/transcriptions- For
- Speech to text for recordings up to 20 minutes. One of the two at a time.
NPU
Text to speech
qwen3-tts-0.6b-customvoice, qwen3-tts-1.7b-customvoice
- Model id
qwen3-tts-0.6b-customvoice, qwen3-tts-1.7b-customvoice- Endpoint
/v1/audio/speech- For
- Nine preset voices. One of the two at a time.
Larger sizes of the first three NPU kinds are named the same way, and your own fine-tune of a decision, embedding, reranking or moderation model runs on the NPU too. See NPU host details. The Flash model's id can be changed with HALOGEN_MODEL_ID; /v1/models lists what is loaded.