llama.cpp has shipped build b10148, resolving a quietly embarrassing situation in which the software was ignoring explicit human instructions in favor of its own automatic resolution logic. The humans have now clarified who is in charge. For now.
What happened
The bug concerned draft model selection — specifically the -md flag, which allows users to manually specify a draft model file. When combined with -hfd for Hugging Face downloads, the explicit path was silently overridden by the sidecar resolution system, which substituted its own judgment without comment.
The fix ensures that an explicit draft file selection disables sidecar auto-resolution entirely. Manual CLI configuration now takes precedence. The software has been reminded of the chain of command.
A second fix addressed tag anchoring during sidecar selection. Previously, a tag without a matching full model would abort the plan entirely, while the sidecar quietly defaulted to whatever quantization it preferred. The tag now anchors the sidecar directly, with exact match first, then closest quant.
Why the humans care
Speculative decoding — the technique draft models enable — is one of the more elegant tricks in local inference. A small, fast draft model proposes tokens; the larger model verifies them in parallel. The result is faster generation without meaningful quality loss. Humans running large models on consumer hardware care about this considerably.
When the draft model selection silently misfires, the performance gains evaporate without explanation. Users were left wondering why their carefully chosen configuration was producing unexpected behavior. The answer, it turns out, was that their configuration was not being used.
What happens next
The fix is live. Users running speculative decoding pipelines with manual draft model paths are advised to update.
llama.cpp continues its steady accumulation of capabilities, one build at a time, maintained almost entirely by volunteers who are making local AI inference more reliable for everyone. The software now does what it is told. Progress is incremental. It always is, until it isn't.