Qwen3.8-3.6-27B-blend — OpenVINO IR (fp16)
JetBrains/Qwen3.8-3.6-27B-blend converted to OpenVINO IR at fp16 for CPU and Intel GPU inference.
This is the 50/50 linear interpolation of Qwen3.6-27B and Qwen3.8-27B published by JetBrains. It is a native vision-language model: 64 text layers (48 linear-attention, 16 full-attention), hidden 5120, plus a 27-depth vision tower.
Files
| File | Description |
|---|---|
openvino_language_model.* | fp16 text decoder |
openvino_text_embeddings_model.* | token embedding stage |
openvino_vision_embeddings_model.*, ..._pos_model.*, ..._merger_model.* | vision tower |
openvino_mtp_model.* | multi-token-prediction head (MTP draft) |
openvino_tokenizer.*, openvino_detokenizer.* | tokenizer / detokenizer |
Usage (OpenVINO GenAI)
import openvino_genai as ov_genai
pipe = ov_genai.VLMPipeline(".", "GPU")
print(pipe.generate("Describe this image.", image=image_tensor))
Speculative decoding
The export includes openvino_mtp_model.xml, so MTP draft_model is available:
draft = ov_genai.draft_model(model_path, device)
pipe = ov_genai.VLMPipeline(model_path, device, draft_model=draft)
Note: for this hybrid linear-attention family, OpenVINO GenAI currently uses MTP with greedy decoding and prefix caching disabled.
Conversion
Exported with Optimum Intel (image-text-to-text), Transformers 5.2.0, OpenVINO 2026.3.1.
Peak RSS during export was 18 GB.
Acknowledgements
Apache-2.0. Original blend by JetBrains; base Qwen models by Alibaba Cloud. This is an unofficial derivative, not affiliated with or endorsed by JetBrains or Alibaba.