Languages & libraries

llama.cpp

Name: llama.cpp
Author: ggml-org

ggml-org

Homepage Repository

b9744 fresh latest release
21.06.2026

Upgrade assessment

Unkritisch

Upgrade unkritisch

Keine offenen Regressionen, kein Breaking, keine aktuellen CVEs

Derived automatically from release, repo and CVE data — no judgment by a language model.

StreamWhat this means for this app

REL b9580 09.06.2026

llama.cpp version b9580 extends Vulkan support with v_dot2_f32_f16 for matrix-matrix multiplication and Flash Attention. Additionally supports Valve FP16 dot2 extension and introduces dot_product abstraction to reduce preprocessor branching. The update contains no breaking changes and can be applied as routine maintenance.

New Features

Vulkan v_dot2_f32_f16 support in matrix-matrix multiplication and Flash Attention
Support for Valve FP16 dot2 extension
dot_product abstraction to reduce preprocessor branching

Read original

REL b9577 09.06.2026

llama.cpp version b9577 introduces the new --log-prompts-dir option, which writes each prompt to a separate text file in a specified directory. This feature facilitates debugging and analysis of inputs. This is a routine update with no breaking changes or security fixes.

New Features

Added --log-prompts-dir option to write each prompt to a separate text file in specified directory

Read original

REL b9575 09.06.2026

llama.cpp version b9575 introduces the new GGML_OP_COL2IM_1D operation for 1D transposed convolution overlap-add steps. The CPU implementation supports F32, F16, and BF16 data types with F32 accumulator. Additionally, backend test coverage and equivalence tests were added, proving that mul_mat + col2im_1d matches native ggml_conv_transpose_1d. For operators, this means extended functionality without breaking changes.

New Features

GGML_OP_COL2IM_1D operation for 1D transposed convolution overlap-add step
CPU implementation supporting F32, F16 and BF16 with F32 accumulator
Backend test coverage for GGML_OP_COL2IM_1D
Equivalence test proving mul_mat + col2im_1d matches native ggml_conv_transpose_1d

Read original

REL b9570 09.06.2026

llama.cpp version b9570 introduces a clang-format job for ggml-webgpu. This change affects only code formatting and the development environment, with no impact on functionality or running operations. No update is required for operators as this is purely a development improvement.

New Features

clang-format job for ggml-webgpu

Read original

REL b9568 08.06.2026

llama.cpp version b9568 adds support for Gemma-4 E2B and E4B assistant variants. The update also introduces masked embedding tensors support for the gemma4-assist architecture. Changes are backward compatible and require no migration of existing deployments. Operators using Gemma-4 models gain extended functionality without service disruption, while other configurations remain unaffected by this routine update.

New Features

Support for Gemma-4 E2B and E4B assistants
Masked embedding tensors support for gemma4-assist architecture

Read original

REL b9565 08.06.2026

llama.cpp version b9565 improves WebGPU support with enhanced buffer overlap handling for the concat operator. This change only affects users running WebGPU as their backend and may improve performance during concatenation operations. Operators using other backends (CPU, CUDA, OpenCL) will see no changes. The update is non-critical and can be applied during routine maintenance.

New Features

WebGPU buffer overlap handling for concat operator

Read original

REL b9564 08.06.2026

llama.cpp version b9564 extends WebGPU support with 2D workgroups implementation for scale, binary, and unary operations. This implementation may improve GPU performance for corresponding computations. The update is backward compatible and requires no configuration changes. Operators can apply this routine update to benefit from the WebGPU optimizations.

New Features

WebGPU 2D workgroups implementation for scale, binary, and unary operations

Read original

REL b9562 08.06.2026

llama.cpp b9562 adds video input capabilities to the framework. The mtmd module now supports video processing, the server accepts Base64-encoded video input, and new CLI parameters like --video enable video handling from the command line. For operators, this means expanded multimodal functionality without breaking existing deployments - current setups continue working unchanged while new video features can be optionally enabled through configuration.

New Features

Video input support in mtmd module
Base64 video input support on server
MTMD_VIDEO configuration option
Video timestamp support
CLI video auto-completion
--video command line argument

Read original

Timelineall entries

Type Ereignis · Quelle Value · Time

REL

Languages & libraries fresh

Release 2026-06-21