From the model card
Fields below are copied from the tags and counters on the HuggingFace repository Qwen/Qwen3.6-35B-A3B at our last fetch. They are set by the uploader, not verified by us; rows with no tag are omitted. How this page is made.
- Publisher (HF namespace)
- Qwen
- Pipeline tag
- image-text-to-text
- Library
- Transformers
- Weight formats
- safetensors
- License tag
apache-2.0— read the license file in the repo before relying on it- Downloads (HF counter at last fetch)
- 4,565,289
- Likes (HF counter at last fetch)
- 2,771
- Model card
- https://huggingface.co/Qwen/Qwen3.6-35B-A3B
Use cases
- Cost-efficient serving of a large-capacity model
- Multimodal reasoning where per-token compute budget is constrained
- Batched inference workloads that benefit from MoE parallelism
- Fine-tuning a high-capacity base without dense 35B memory cost
Pros
- 3B active params means inference cost similar to a 3B dense model
- 35B total capacity stores more knowledge than small dense models
- Apache-2.0 licensed
- Strong multilingual support from Alibaba's training corpus
Cons
- All 35B parameters must fit in memory even if only 3B active per token
- MoE models exhibit expert routing instability during fine-tuning
- Less community fine-tune coverage than Mistral or Llama MoE variants
- FP16 requires ~70GB VRAM across devices
Tags
transformerssafetensorsqwen3_5_moeimage-text-to-textconversationallicense:apache-2.0eval-resultsendpoints_compatibledeploy:sagemakerdeploy:azureregion:us