Unsloth Fixes Mistral Medium 3.5 GGUFs

This is a reminder that for local LLMs, the quantization and conversion layer can matter as much as the base weights. A strong model becomes a bad one fast if the GGUF packaging is off.

• –The bug hit the exact use case local users care about most: long-context reliability, where small conversion errors become obvious
• –The fix suggests the base Mistral model was not the problem; the packaging/conversion pipeline was
• –Vision support still looks unfinished, so this is a correctness fix, not a full feature stabilization story
• –For teams running self-hosted inference, this is a cautionary tale to re-test every new quant and not assume upstream weights are enough
• –The quick turnaround is good news for the local LLM ecosystem, where Unsloth and llama.cpp compatibility directly affect adoption