llama.cpp b12180 adds text-only Kimi-K3.1-Pro support… Vulkan detects ReBAR
Posted As of
On November 6, llama.cpp build b12180 added text-only Kimi-K3.1-Pro support. The same day, the Vulkan backend began detecting whether ReBAR is supported.
What happened
Based on the llama.cpp release list (b12171 to b12182, November 6-7).
- b12180 (November 6): Text-only Kimi-K3.1-Pro model support. It reuses the K3 graph and converter, puts a lightning indexer in 24 of the 93 layers, and uses a new hybrid cache. (official)
- b12171 (November 6): The Vulkan backend detects whether ReBAR is supported and avoids BAR allocation if it is not. (official)
- b12173 (November 6): The repeat operation in the OpenCL backend supports f16, i32, i16, and bf16. (official)
- b12177 (November 6): Fixed a bug where the CLI crashed on invalid UTF-8 in prompts or files. (official)
What the release notes do not say: required memory, actual inference speed, or whether this model runs on an ordinary PC. Support being added and the model being practically usable are different things.
Sources and related posts
- llama.cpp GitHub 릴리스 노트 Primary source