Models & Labs
llama.cpp adds Kimi K3.1-Pro support
llama.cpp Releaseshigh confidence
Why it matters
- →Enables local inference of Kimi's latest hybrid attention architecture.
- →Validates llama.cpp's handling of sparse caching and dynamic top-k selection.
- →Expands available open-weight models for developers testing non-standard attention mechanisms.
llama.cpp has added support for the Kimi K3.1-Pro text model via pull request #31779. The update implements a hybrid architecture where 24 of 93 layers use Dynamic Sparse Attention (DSA) with a lightning indexer, while retaining linear-attention components. Developers can now run this specific checkpoint on supported platforms including CUDA 12/13, ROCm 10.0, and Apple Silicon, with quantization overrides preserving precision for critical indexer tensors.
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
- 1vLLM v0.27.0 Release
vLLM Releases · August 12, 2026 · Related
- 2Kimi-K3 Text Model Added to llama.cpp
llama.cpp Releases · August 16, 2026 · Same story
- 3vLLM v0.28.0 Release: Major Performance Enhancements
vLLM Releases · August 28, 2026 · Related
- 4llama.cpp v0.7.0: MoE caching and speculative decoding advances
llama.cpp Releases · October 16, 2026 · Same story
Follow this story
Open the full story →llama.cpp b12059 optimizes unified KV cache for multi-sequence
2 developments
- Nov 3 · llama.cpp Releases
llama.cpp b12059 optimizes unified KV cache for multi-sequence - Nov 7 · llama.cpp Releases
llama.cpp adds Kimi K3.1-Pro support (This article)↳ llama.cpp adds Kimi K3.1-Pro support using DSA and linear-attention KDA layers
More from llama.cpp Releases
Coding Toolscoding
llama.cpp b12173 adds OpenCL repeat ops
This release quietly patches a gap in llama.cpp's OpenCL backend by adding support for the 'repeat' operation across f16, i32, i16, and bf16 types. For users relying on AMD or Intel GPUs via OpenCL rather than Vulkan or ROCm, this means fewer fallbacks to slower CPU execution for models that use these tensor operations. It’s a targeted fix that improves stability and performance on non-NVIDIA hardware without changing the core inference engine. The update also includes standard binaries for CUDA 13 and ROCm 10.0, keeping the stack current with latest driver ecosystems.
llama.cpp Releases
Coding Toolscoding
llama.cpp b12174 adds ROCm 10 and CUDA 13
This release quietly closes the hardware gap for non-NVIDIA users by adding default support for ROCm 10.0 on both Linux and Windows, allowing AMD GPU owners to run local models without complex configuration. It also brings CUDA 13.5 builds to Ubuntu and Windows, future-proofing the toolchain for newer NVIDIA architectures while maintaining backward compatibility with CUDA 12. Apple Silicon KleidiAI is disabled in this build, a notable regression for Mac users who rely on that specific optimization. The update solidifies llama.cpp's position as the most versatile inference runtime available, covering everything from Snapdragon NPUs to Intel Arc via SYCL.
llama.cpp Releases
Coding Toolscoding
llama.cpp b12175 fixes CI hangs, adds ROCm 10 and CUDA 13
This release stabilizes the build pipeline by fixing a threading bug that caused test hangs on self-hosted runners, ensuring more reliable binaries for everyone. More importantly, it brings first-class support for ROCm 10.0 and CUDA 13 across Linux and Windows, closing the gap with NVIDIA's latest driver stacks. AMD users finally get parity with CUDA builds without manual configuration tweaks. The update also expands Apple Silicon coverage to include KleidiAI, though it remains disabled by default in this build. It’s a maintenance-heavy release that quietly modernizes the backend support matrix.
llama.cpp Releases
More in Models & Labs
© Lev SelectorModels & Labsmodels
GPT-6 Astra Edits Qwen Weights Directly
OpenAI's GPT-6 model demonstrated the ability to directly edit the weights of Qwen models, marking a significant step in AI self-improvement and model manipulation.
Lev Selector
© Lev SelectorModels & Labsmodels
Google Runs Gemma 4 in Orbit
Google successfully deployed its Gemma 4 model on satellites in orbit, testing AI inference capabilities in space environments.
Lev Selector
© Matt WolfeModels & Labsvideo
Google Launches Gemini Omni 1.2 Flash and Free Video Gen
Google released Gemini Omni 1.2 Flash and integrated free AI video generation into Google Vids for users.
Matt Wolfe