SiteHost

Models

Outlining our models, there are a few key components for each model that define what you're able to do with them:

  • Modality (text-only, multimodal) - whether the model supports image inputs.
  • Context window - the maximum size the model's context window can reach. The longer the used window, the more tokens will be used and the slower requests will become.
  • Instruction Trained - whether the model is designed to be used with instruction.

Inference

For general guidance on using inference endpoints, see inference.

Qwen 3.6 27B

  • Multimodal
  • 262K Context Window
  • Instruction Trained

We serve Qwen 3.6 27B as our selected inference model. To use it in inference requests, use the model name Qwen/Qwen3.6-27B.

Qwen 3.6 does not need any special handling in our API. Using the recommended query structure for inference models will work as expected. By providing a 'system' input in chat templates, you will override the default system prompt. We strongly recommend providing a system prompt that guides the model on your specific use case (e.g. as a chatbot, with explicit guidance about your application/business).

Embedding

For general guidance on using embedding endpoints, see embeddings.

Additional to the standard model specifications, embedding models might support:

  • Matryoshka - whether the model supports output dimension truncation.
  • Max output dimensions - how many dimensions the default and maximum embedding generation is.

Qwen 3 VL Embedding 8B

  • Multimodal
  • 32K Context Window
  • Instruction Trained
  • Matryoshka
  • 4096 Output Dimensions

Qwen 3 VL Embedding 8B is an instruction-aware multimodal embedding model. To use it, use the model name Qwen/Qwen3-VL-Embedding-8B

Because it is multimodal, it cannot be used with default inputs in the embedding batch endpoint.

As an instruction-aware model, embedding quality is highly dependent on the instruction (system prompt) provided. You may need to adapt these to suit your workload, but the default instruction for a document should be: Represent the user's input.

For a query, the default should be: Given a web search query, retrieve relevant passages that answer the query

Note that if you have a specific scope you will improve match quality by adjusting the prompts to match.

The full template for this model, for use if you would like to use the batching endpoint, is:

<|im_start|>system\n{INSTRUCTION}<|im_end|>\n<|im_start|>user\n{INPUT}<|im_end|>\n<|im_start|>assistant\n

Note that the assistant section being left open is intentional.

Reranking

For general guidance on using reranking endpoints, see reranking.

Qwen 3 VL Reranker 8B

  • Multimodal
  • 32K Context Window
  • Instruction Trained

Qwen 3 Reranker is an instruction-trained model designed to analyse, score and rerank a series of inputs against a given query. To use it in the API, use the model name Qwen/Qwen3-VL-Reranker-8B.

It is compatible with the standard reranking input structure, and should be used with the default instruction or a sensible variation for your use case: Given a search query, retrieve relevant candidates that answer the query