vLLM GGUF support still looks experimental

Interest in this question is a good signal that developers want vLLM’s serving stack to handle the same cheap, portable quantized models they already use elsewhere. But today GGUF in vLLM still reads like a compatibility path, not a production-default format.

• –vLLM’s official GGUF page explicitly warns that support is “highly experimental and under-optimized” and may be incompatible with other features
• –Current docs say only single-file GGUF models are supported, so multi-part GGUF checkpoints have to be merged before use
• –The project recommends using the base model tokenizer because GGUF tokenizer conversion is slow and unstable on some models
• –Community discussion around 2025–2026 still regularly frames GGUF-in-vLLM as slower and rougher than GGUF-first stacks like llama.cpp
• –The practical takeaway for infra teams is simple: use vLLM if you want its serving engine and OpenAI-style API, but don’t assume GGUF is the mature fast path yet