JetBrains/Qwen3.8-3.6-27B-blend
The JetBrains/Qwen3.8-3.6-27B-blend is an unofficial JetBrains derivative model, created by linearly interpolating the checkpoint parameters of Qwen3.6-27B and Qwen3.8-27B. This 27 billion parameter model, with a 32768 token context length, is optimized for coding tasks, as demonstrated by its performance on JetBrains' internal coding benchmark. It retains the configuration, tokenizer, processor, and chat template from Qwen3.8-27B, offering a blend of capabilities from its Qwen predecessors.
Loading preview...
Overview
JetBrains/Qwen3.8-3.6-27B-blend is an unofficial derivative model developed by JetBrains, combining the strengths of Qwen3.6-27B and Qwen3.8-27B. This 27 billion parameter model was created by a 50/50 linear interpolation of their checkpoint parameters, using float32 accumulation. It leverages the configuration, tokenizer, processor, and chat template of Qwen3.8-27B.
Key Characteristics
- Blended Architecture: A direct interpolation of two Qwen models, aiming to combine their respective capabilities.
- Coding Performance: Evaluated on JetBrains’ internal 100-task coding benchmark, demonstrating its proficiency in code generation and reasoning. Further details on methodology and results are available in the Making Local AI Smarter and Faster article.
- No Additional Training: This derivative was created solely through checkpoint merging; no new training or fine-tuning was performed, and no additional data was used.
- Multiple Formats: Available in BF16, GGUF, MLX 4-bit, and MTP MLX 4-bit formats for diverse deployment needs.
Licensing and Attribution
Distributed under the Apache License, Version 2.0, retaining the original Qwen license and copyright by Alibaba Cloud. JetBrains holds copyright for its original contributions to this derivative.