MiMo-V2.5 GGUF preview lands

This is the practical layer that makes a big base model usable locally, but it is still early and tied to upstream inference work.

• –The quant repo targets llama.cpp directly, so the real milestone is deployment readiness rather than a new model architecture
• –The MoE-aware quant scheme is the interesting part: keep core weights high quality while compressing FFN-heavy tensors harder
• –The text-only limitation matters because MiMo-V2.5’s native multimodal abilities are not exposed in this GGUF yet
• –Pre-merge support means anyone adopting it early should expect churn in weights, conversion scripts, or runtime behavior
• –This is a strong signal that the local-LLM ecosystem will get rapid third-party coverage once the upstream PR stabilizes