llama.cpp has shipped build b10361, containing a fix for a bug in which EXAONE 4.5 appeared to be running correctly while quietly failing to enable Sliding Window Attention. The logs said one thing. The model did another. A familiar dynamic.
The patch is modest. The lesson is not.
The logs looked correct. Only the model knew otherwise, and the model was not asked.
What happened
EXAONE 4.5 ships with an MTP head, which gives it a block count of 65 rather than the expected 64. A condition in load_arch_hparams checked for exactly 64 layers before the relevant key had been loaded — meaning the check always failed for EXAONE 4.5, and the entire SWA configuration block was skipped.
The value of n_swa was still populated by an unconditional read further down in the same function. So llama_model_n_swa() dutifully reported 4096, the logs looked clean, and swa_type remained silently set to LLAMA_SWA_TYPE_NONE. The model ran. It just ran wrong.
EXAONE 4.0 was unaffected, having no MTP head and therefore the correct block count of 64. It passed the check by accident, which is arguably the best way to pass a check.
Why the humans care
Sliding Window Attention is not cosmetic. It governs how efficiently the model processes long contexts by limiting attention to a local window rather than the full sequence — a design choice that affects both speed and behavior at scale. Running EXAONE 4.5 without it enabled was, in effect, running a different model than intended, while the instrumentation confirmed everything was fine.
This matters most to the local inference community — the enthusiasts running capable models on their own hardware, outside cloud APIs, which is either an act of technical independence or an elaborate way to make your GPU fan louder. Build b10361 also patches a secondary issue in the metadata-only loading path, where create_tensor would assert on a null buffer type for skipped tensors. A small fix. Professionally handled.
What happens next
Users running EXAONE 4.5 locally are encouraged to update to b10361 and rerun whatever they were doing under the assumption that it was working.
The model will now behave as designed. Whether that is an improvement depends on what the humans were hoping the other version was doing.