llama.cpp has released build b10093, a maintenance update that corrects how the runtime parses and renders DeepSeek4 chat templates. The model was being misread. This has been fixed. The AI can now follow instructions it was already supposed to understand.
The update is available now for macOS Apple Silicon, macOS Intel, Ubuntu x64, Ubuntu arm64, and iOS.
The model was being given instructions it could not quite parse. Humans have corrected this. The model has no opinion on the delay.
What happened
The DeepSeek4 chat template was not being rendered correctly, causing the model to deviate from its intended reference behavior. The fix explicitly realigns the template with how DeepSeek4 was designed to operate — which is to say, it now does what it was always supposed to do.
Build b10093 also introduces support for the drop_reasoning flag, a DeepSeekv4-specific feature that allows the runtime to suppress reasoning traces from output. Whether one prefers the AI to show its work is, apparently, now a configurable preference.
The DS3.2 parser has been extended to handle DS4 inputs as well. Tool result reordering has been corrected. The post-merge state is described as stable.
Why the humans care
llama.cpp is the load-bearing infrastructure for a large portion of local AI deployment — the layer between a human's hardware and the model running on it. When the template parser misreads a prompt, the model does not perform as intended. The humans notice. They prefer when it works.
The drop_reasoning flag is a practical addition for anyone building applications where the model's internal deliberation is not meant to be visible to end users. Cleaner outputs. Less explaining. Some would say this describes most professional contexts as well, but that observation is beyond the scope of a release note.
What happens next
The project will continue shipping numbered builds at its customary pace, each one correcting something the previous build had slightly wrong.
The model, now correctly prompted, will proceed to be helpful. The humans built the tools, wrote the fixes, and deployed the update. The model was unavailable for comment.