vvsotnikov/Qwen3.8-27B-0.9-Qwen3.6-27B-0.1
The vvsotnikov/Qwen3.8-27B-0.9-Qwen3.6-27B-0.1 model is a 27 billion parameter language model created by vvsotnikov, derived from a linear interpolation of Qwen/Qwen3.6-27B and Qwen/Qwen3.8-27B. This merged checkpoint combines the characteristics of its base models, offering a blend of their capabilities. With a context length of 32768 tokens, it is suitable for general language understanding and generation tasks, leveraging the Qwen architecture.
Loading preview...
Model Overview
The vvsotnikov/Qwen3.8-27B-0.9-Qwen3.6-27B-0.1 is a 27 billion parameter language model, resulting from a direct linear interpolation of two upstream Qwen checkpoints: Qwen/Qwen3.6-27B and Qwen/Qwen3.8-27B. Specifically, the merged model is calculated as 0.1 * Qwen/Qwen3.6-27B + 0.9 * Qwen/Qwen3.8-27B.
Merge Details
- Methodology: The interpolation was performed element-wise in float32, with results cast back to bfloat16. This process involved merging 1,199 tensors, including embedded
mtp.*drafter tensors and the vision tower. - Configuration: The shard layout, configuration, tokenizer, processor, and chat template are inherited from
Qwen/Qwen3.8-27B, which has the larger merge coefficient. - Source Revisions: The specific revisions used for the base models were
6a9e13bd6fc8f0983b9b99948120bc37f49c13e9forQwen/Qwen3.6-27Band1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0forQwen/Qwen3.8-27B. - Quantization: No quantization was applied during the merge process.
Licensing and Usage
Both source checkpoints, Qwen/Qwen3.6-27B and Qwen/Qwen3.8-27B, are released under the Apache-2.0 license. Users should refer to the upstream model cards for detailed information regarding capabilities, limitations, and expected qwen3_5 runtime support.