Mistral Medium 3.5 128B struggles in llama.cpp

This looks more like a local inference / quantization mismatch report than evidence that the base model is fundamentally broken.

• –The thread centers on llama.cpp, Vulkan, and Q4 quants, so the observed weakness could come from backend support, quantization quality, or prompt/sampling settings rather than the model weights themselves.
• –The user is comparing it against Magistral Small and finding worse coding and reasoning performance, which suggests a configuration or compatibility issue worth cross-checking in other runtimes like vLLM.
• –Since Mistral positions Medium 3.5 as a 128B dense open-weight flagship for coding and agentic work, reports like this are relevant because they may indicate that some local stacks are not yet getting the intended behavior.
• –As a community signal, it is useful but not conclusive: one Reddit thread is enough to flag a possible issue, not to establish a model-wide regression.