QwenMix-3.7: Community merges Qwen3.6 and 3.8 into working hybrid
A Reddit user combined Qwen3.8 and 3.6 because they're nearly identical — and it works.
Deep Dive
Reddit user bigattichouse decided to try something silly: combining Qwen3.8-27B and 3.6-27B using the Qwen3.8-27B-UD-Q6_K_XL.gguf file—even though 3.8 has seven new tokens. It sorta works, but only after a smoke test. Scripts and the idea are in the replicate/ folder inside the model repo. Might prove useful to someone.
Key Points
- QwenMix-3.7 merges Qwen3.8-27B and Qwen3.6-27B using the Qwen3.8-27B-UD-Q6_K_XL.gguf file
- Qwen3.8 only differs from Qwen3.6 by seven new tokens, making the merge straightforward
- The model passes smoke tests, but the creator stresses no deep testing was done yet
Why It Matters
Shows how close Qwen3.6 and 3.8 really are, and opens the door for community-driven model merging.