Webs-KoReasoner-27B-v1

Recipe

• Base: Qwen/Qwen3.5-27B (Apache-2.0)
• Donors (same base):
  • NewenAI/QuettaLLMs-27B-Koreasoner-V3
  • jiwon9703/Qwen3.5-KoReasoin-27B-v1
• Method: DARE-TIES (density 0.5, standard 1/density rescale, weight 1.0, seed 42), TIES sign-election across the two task vectors, applied to all tensors (including MLP). fp32 math, streamed one tensor at a time so a 64GB machine never holds more than one tensor's working set.

Because mergekit cannot express the Qwen3_5 hybrid architecture (interleaved linear_attention / full_attention layers), the merge was produced with an in-house streaming merger.

Intended use

Korean knowledge & reasoning. The model thinks (often in English) inside <think> … </think> and answers in Korean.

Credits

All weights derive from the Apache-2.0 base and donors above; full credit to their authors. Merge engineering by 웹스팩토리 (Websfactory).