User reports MiMo-V2.5 looping issues with unsloth ud-q4_k_xl quant

A Reddit user reports that the MiMo-V2.5 model exhibits persistent reasoning loops when used with OpenCode and the unsloth ud-q4_k_xl quantization. The user notes that while the reasoning spans are legitimate, the model struggles to make decisions without manual intervention.

• The user compares MiMo-V2.5 favorably against Qwen 3.5 397B, citing better web search capabilities and fewer hallucinations.
• Qwen 3.5 397B reportedly hallucinated a plan using both Vulkan and DX12 simultaneously with a made-up Vulkan version.
• MiMo-V2.5 successfully scrapped the flawed plan and researched alternatives using only webfetch.
• The user has approximately 30GB of VRAM remaining and considers increasing the quantization quality to see if it resolves the looping issue.