Models & Labs
llama.cpp b9952 Release Enhances DeepSeek V4
llama.cpp Releaseshigh confidence
Why it matters
- →The update improves model efficiency by optimizing KQ masks and removing biases.
- →Streamlined code enhances performance and maintainability.
- →Supports a wide range of platforms, increasing accessibility for developers.
The b9952 release of llama.cpp introduces several enhancements aimed at improving the DeepSeek V4 model. Key updates include converting KQ masks to f16 when FA is used and eliminating zero attention bias, which together enhance model efficiency. The update also removes unnecessary raw_k repeats, simplifying the code. These changes are designed to optimize performance across multiple platforms, including macOS, Linux, and Windows, making llama.cpp a more efficient tool for AI developers.
Read originalThe story around this
Earlier coverage that leads up to this article, and what followed. Lines connect each piece to the closest one after it, converging here.
How we got here
- 1Llama.cpp b9840 Release Enhances DeepSeek V4
llama.cpp Releases · June 30, 2026 · Same story
What happened next
- 2Llama.cpp b10093 Release Fixes DeepSeek4 Template
llama.cpp Releases · July 25, 2026 · Same story
More from llama.cpp Releases
Coding Toolscoding
llama.cpp b12022 adds ROCm 10 and Snapdragon support
This release quietly cements llama.cpp as the universal inference runtime by finally bringing ROCm 10.0 to Linux and Windows alongside CUDA 13.4, effectively closing the hardware gap for AMD users who were previously left behind. The inclusion of native Snapdragon support on Linux marks a significant step toward running local LLMs on ARM-based mobile chips without emulation overhead. While KleidiAI on Apple Silicon is disabled in this build, the broader expansion to diverse accelerators ensures that developers aren't locked into NVIDIA ecosystems anymore.
llama.cpp Releases
Coding Toolscoding
llama.cpp b12023 fixes shared prefix bug
This release addresses a specific inference artifact where decision prefixes were incorrectly bleeding into output embeddings. By capping shared prefixes at the first output position, llama.cpp ensures that option tokens remain distinct for parent and child tasks. This is a targeted fix for multi-step reasoning or voting workflows rather than a performance boost. It stabilizes complex generation patterns without changing the underlying model architecture.
