llama.cpp b10328 adds Docker-based tool isolation to its server, letting local LLM function calls run in containers instead of on your host machine.
llama.cpp build b10331 fixes a get_info bug that returned the wrong working directory when tools ran in isolates. KleidiAI macOS build disabled.
llama.cpp b10306 adds a contiguous fast path for fused GLU ops on SYCL, with +14% f16 SWIGLU gains on Intel Arc Pro B70. Fused-GLU kernels consolidated.
llama.cpp build b10321 fixes a Metal kernel bug where partial simdgroups caused incorrect NORM/RMS_NORM outputs on Apple Silicon. No crash — silently wrong results.
llama.cpp b10322 coalesces SSM convolution memory loads on Intel SYCL, delivering up to 1.87x kernel speedup on Arc Pro B70. End-to-end gains on Qwen 3.5 27B are modest at +2.2%.
llama.cpp release b10295 ships a single fix for quantized reshaped tensor strides in the model loader (PR #26672), with builds across macOS, Linux, Windows, Android, and iOS.
llama.cpp build 10297 patches an empty-response bug in the /cors-proxy endpoint and ships CUDA 13.3 DLLs, but drops macOS Intel and KleidiAI builds.
llama.cpp build b10282 adds spec-decode counters to the server /metrics endpoint with vLLM-aligned parameter names, plus CUDA 13.3 build artifacts.
llama.cpp build b10285 ships multi-row batching for DeepSeek OCR, processing image rows in one pass. Pre-built binaries span 25+ platforms.
Build b10286 of llama.cpp ships a grammar change that silently switches max repetition values of 2000+ to unbounded, plus CUDA 13.3 DLLs for Windows.
llama.cpp build b10289 fixes three bugs in server file_glob_search: Windows junction loops, case-insensitive junk dirs, and silent directory failures.
llama.cpp build b10251 adds Multi-Token Prediction for GLM-4.7-Flash and ships prebuilt binaries across macOS, Linux, Windows, Android, and openEuler.
llama.cpp b10255 extends Intel GPU oneDNN attention to Q4_0–Q8_0 and FP32 KV caches. Claude co-authored the PR. BF16 and IQ types excluded.
llama.cpp build b10259 adds tensor reshape during model load and ships CUDA 13.3 DLLs. Full platform matrix across macOS, Linux, Windows, Android.
llama.cpp release b10270 adds Qwen3-TTS to its mtmd pipeline and ships a breaking change to the llama-tts binary. Here's what changes for local TTS users.
llama.cpp build b10271 adds per-conversation CWD for agent mode, with path navigation, persistent tool CWD, and UI pickers. Claude assisted.
Build b10275 patches a bug where Windows OEM code pages silently dropped accented characters from child process output in llama.cpp server built-in tools.
llama.cpp build b10237 adds Multi-Token Prediction support for DeepSeek V3.2, fixes model type discovery, and ships pre-built binaries across major platforms.
llama.cpp build b10238 adds Multi-Token Prediction support for Qwen3-Next, with prebuilt binaries across CPU, Vulkan, ROCm, OpenVINO, and SYCL backends.
llama.cpp b10224 ships f16 repeat for the ggml-webgpu backend and prebuilt CUDA 13 DLLs across 30+ targets including ROCm 7.2, SYCL, and OpenCL Adreno.
llama.cpp b10231 adds DSpark speculative sidecars with new resolution rules. DSpark outranks dflash in auto-selection. Prebuilts cover 20+ platform variants.
Build b10234 adds float16 support for binary operations on Apple Metal, plus prebuilt CUDA 13.3 and ROCm 7.2 binaries across 25+ configurations.
Build b10217 enables tool call parsing during DeepSeek reasoning phases, plus CUDA 13.3 and ROCm 7.2 Windows/Linux builds.
llama.cpp b10219 fixes reasoning_content persistence in chat history, enabling --reasoning-preserve to work across multi-turn conversations with local reasoning models.
llama.cpp build b9994 adds Q2_0 quantization support to the Metal backend, enabling 2-bit models on Apple Silicon with lower memory use.
llama.cpp b9995 doubles SYCL flash attention threads to 256 for Intel Battlemage xe2 GPUs, keeps 128 for ARC Alchemist, adds CUDA 13.3 and ROCm 7.2 builds.
llama.cpp release b9990 adds EAGLE3 speculative decoding for Minimax2 models, ships CUDA 13.3 prebuilt DLLs, and covers Ascend NPU builds via openEuler.
llama.cpp b9992 refactors CUDA MMQ kernels, fixes Blackwell for RTX 50-series, adds CUDA 13.3 and ROCm 7.2 builds, disables macOS Intel.
llama.cpp b9968 adds int8 dp4 dense and MoE GEMM OpenCL kernels for Qualcomm Adreno GPUs. Build artifacts span 20+ platform configs including Windows arm64 OpenCL.
llama.cpp release b9969 routes large matmuls to medium tile on Adreno, fixing llama-cli crashes on longer prompts with q4_0 quantized models.
llama.cpp build b9974 fixes a CUDA crash where querying memory on a GPU with no free VRAM killed the entire inference process before it started.
llama.cpp b9979 fixes a silent prompt truncation bug where embedded NUL bytes dropped messages with no log. PR #25548 adds text_len to mtmd_input_text.
llama.cpp b9965 ships bitonic sort in HVX registers for Hexagon ARGSORT up to 1024 elements, plus CUDA 13.3 DLLs and a 20+ platform build matrix.
llama.cpp b9966 fixes a decode-thread bottleneck: 29 std::regex patterns recompiling per tensor per token in -sm mode are now static const. PR #24710.
Release b9951 adds a new ggml ET backend to llama.cpp, shipping MUL_MAT, ROPE, RMS_NORM, and 8 other compute kernels for local LLM inference.
llama.cpp build b9957 reworks the server tool system: apply_diff removed, tools_io abstraction added, edit tool improved. Prebuilt binaries across 30+ platforms.
llama.cpp build b9934 tunes WebGPU subgroup split in flash attention and ships prebuilt binaries for 30+ configs including CUDA 13.3 and ROCm 7.2.
Release b9938 enables -funsafe-math-optimizations for AMD HIP, matching CUDA's fast-math compilation without the NaN issues -ffast-math would cause.
llama.cpp build b9939 adds CONV_2D_DW depthwise convolution to the Metal backend, with F16 CPU fixes and new test coverage across the release.
llama.cpp build b9931 adds two OpenCL optimizations for MoE prefill GEMM that skip padding tiles, reducing wasted compute with byte-identical output.