llama.cpp Releases
609 releases of ggml-org/llama.cpp
- v1.0.0
llama.cpp v1.0.0 is a milestone release that prunes the deprecated public API and promotes the NextN/speculative decoding API into llama.h, alongside major inference features such as OCP FP8 KV caches, a host-memory MoE expert cache with a configurable CPU share, and external PLE/engram shard loading. It adds several new text and multimodal architectures (Kimi-K3.1-Pro, Ripple, xing4_0, drama-3 TTS, MageVL-Cowork, Kimi-K3.1 vision), hardens llama-server against SSRF/CORS/DNS-rebinding attacks, introduces an OpenAI-compatible /v1/audio/speech endpoint, and ships a redesigned web UI. The bundled ggml library is also bumped to v1.0.0 with broad cross-backend performance work.
Nov 13, 2026
- v0.9.0
llama.cpp v0.9.0 introduces OCP FP8 E4M3 quantization, a new flash_attn_ext_rows operation for efficient unified KV cache use in multi-sequence inference, and a generic batched tensor copy API. It adds four new model architectures, expands audio generation and transcription support, and delivers major Gemma 4 prefill optimizations alongside the ggml v0.29.0 update.
Nov 4, 2026
- v0.8.0
v0.8.0 is a large release that introduces the LiquidAI LFM3 family (text, vision, audio and omni) alongside a new experimental streaming-audio API in mtmd and realtime audio support in the server via the new /v1/realtime and /tts endpoints. It adds many new architectures (DeepSeek V4.1 Flash, SenseNova U1, Beam, Kolibri-1, Flint, Muse Spark, Granite Switch Decision, plus decision/embedding models), reworks sparse attention onto a fused masking op, adds a 50/50 SLRU cache for host-resident MoE experts, and improves router memory management, tool calling and the web UI. ggml is bumped to v0.28.0, bringing GGML_OP_TOP_K_MASK, -1 expert-index semantics and substantial backend performance work across CUDA, SYCL, OpenCL, Hexagon, Vulkan and KleidiAI.
Oct 28, 2026
- v0.7.0
llama.cpp v0.7.0 introduces speculative decoding techniques like LiliCorr and MTP for GLM5-Next, along with a new GPU cache for MoE experts that offers significant speedups. The release expands model support with new architectures including K2 Horizon and EmbeddingGemma-2, while updating the server API and adding a comprehensive models manager to the web UI.
Oct 15, 2026
- v0.6.0
llama.cpp v0.6.0 introduces a new extended batch API for mixed token/embedding inputs, adds support for the GLM-5.3-Flash (320B) and Clef decision models, and implements a new `/v1/systemone` API. It also includes Metal flash attention optimizations and a Web UI overhaul.
Oct 5, 2026
- v0.5.0
This release prioritizes backend performance and correctness, introducing support for new models like HRM-Text and MiMo-V2.6, and significantly improving server robustness with multi-address binding capabilities.
Sep 23, 2026
- v0.4.1
llama.cpp v0.4.1 adds support for new models like Maple 20B-A1B and Tencent Hy 4, significantly improves JSON schema handling and logging, and updates the ggml library to version 0.24.0.
Sep 14, 2026
- v0.4.0
llama.cpp 0.4.0 introduces support for Qwen3.8-Flash-Next and Nemotron-3-Puzzle models, adds lazy tensor reading and per-slot server context limits, and updates ggml to 0.23.0 with sparse flash attention and RDMA capabilities.
Sep 4, 2026
- v0.3.0
Major version bump adding dots3-note multimodal model support, GLM-4.5-Air MTP support, and ggml v0.22.0 updates including meta-backend tensor split and Metal parallel compilation.
Aug 25, 2026
- v0.2.0
Major version bump introducing KleidiAI SME2 support, SYCL kernel updates, UI refactor, and deep support for DeepSeek 4 models.
Aug 21, 2026
- b10549
Enables tensor split support for LFM2 and LFM2MOE models in Tensor Parallelism (TP) mode.
Aug 21, 2026
- b10545
Fixes a critical bug in the Metal Tensor API mat-mat kernel where reading out-of-bounds memory caused NaNs when the K dimension was not a multiple of 32.
Aug 21, 2026
- b10541
Adds a command-line argument `--mmproj-device` to the MTMD (Multi-Modal Text-to-Model) backend to allow specifying the device for the multimodal projection model.
Aug 21, 2026
- b10539
Fixes a potential overflow in the Vulkan Flash Attention (FA) MMQ kernel where Q quantization calculations used denormalized floats.
Aug 21, 2026
- b10538
Optimizes the macOS Metal backend by dequantizing the Key-Value (KV) cache only for large batches to reduce overhead and improve efficiency.
Aug 21, 2026
- b10537
Updates the CI build process to use LLVM's OpenMP library instead of the Microsoft debug version, improving compatibility and packaging.
Aug 21, 2026
- b10536
Adds a lazy-loading feature to the server router that delays the loading of startup models until after the main server setup is complete.
Aug 21, 2026
- b10534
Optimizes CUDA decoding performance by tuning the crossover point between mul_mat_vec_q and MMQ batch kernels, adding hardware-specific switch points to handle different GPUs and quantization types.
Aug 21, 2026
- b10533
Improves robustness in JSON schema parsing by gracefully falling back when unsupported regex patterns are encountered, preventing application crashes.
Aug 21, 2026
- b10532
Optimizes Flash Attention on the Metal backend by dequantizing KV cache to F16 in a preprocessing pass, improving accuracy and performance for various quantized KV cache types.
Aug 21, 2026