Method: element-wise linear interpolation in float32, cast to the original bfloat16 dtype after each tensor is merged
Tensors merged: 1,199, including the embedded mtp.* drafter tensors and vision tower
Shard layout, configuration, tokenizer, processor, and chat template: Qwen/Qwen3.8-27B
Qwen/Qwen3.6-27B revision: 6a9e13bd6fc8f0983b9b99948120bc37f49c13e9
Qwen/Qwen3.8-27B revision: 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0
Quantization: none

The two source checkpoints have identical tensor names, shapes, dtypes, and model configuration; their configs differ only in the recorded Transformers version. The newer Qwen/Qwen3.8-27B non-weight files are used because it has the larger merge coefficient.

License and use

Both source checkpoints are Apache-2.0. Their upstream model cards describe capabilities, limitations, and the expected qwen3_5 runtime support.

creekhop/Qwen3.8-27B-0.9-Qwen3.6-27B-0.1-GGUF

Qwen3.8-27B-0.9-Qwen3.6-27B-0.1-GGUF

Source Model Card

Qwen3.8-27B 0.9 + Qwen3.6-27B 0.1

Merge details

License and use

creekhop/
Qwen3.8-27B-0.9-Qwen3.6-27B-0.1-GGUF