The weights are published in the standard Hugging Face Transformers format as 14 Safetensors shards with the accompanying config.json, tokenizer files, processor configuration, and shard index. The model configuration retains the Qwen-compatible chat template supplied in tokenizer_config.json.

Model overview

Property	Value
Base model	`Qwen/Qwen3.5-35B-A3B-Base`
Checkpoint	Together AI Final Merged, step 40
Architecture	`Qwen3_5MoeForConditionalGeneration`
Format	Hugging Face Transformers / Safetensors
Pipeline	Image-text-to-text

Loading with Transformers

python

from transformers import AutoProcessor, Qwen3_5MoeForConditionalGeneration

model_id = "RoseG/Qwen3.5-35B-A3B-Base-8be64d30-Merged"
processor = AutoProcessor.from_pretrained(model_id)
model = Qwen3_5MoeForConditionalGeneration.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)

License

This checkpoint follows the Apache-2.0 license of its base model. See the base model repository for its model documentation and license information.