GGML
PulseAugur coverage of GGML — every cluster mentioning GGML across labs, papers, and developer communities, ranked by signal.
- developed by GGUF 90%
- used by central processing unit 90%
- instance of central processing unit 90%
- founded Georgi Gerganov 90%
- used by audio.cpp 90%
- used by Vulkan 70%
- used by macOS 70%
- developed by audio.cpp 70%
- used by Georgi Gerganov 70%
- used by WebGPU 70%
- used by graphics processing unit 70%
- used by Rocm 70%
16 day(s) with sentiment data
-
New C++ Tool Enables Audio Embedding Inference with GGML Models
A new C++ implementation called clap.cpp has been developed for audio embedding inference using GGML models. This tool supports various audio models like CLAP, CLAP2, CLaMP3, MERT, MuQ, and MuQ-MuLan, and can process WA…
-
DeepSeek-V4-Flash model runs on 64GB CPU via llama-modchip
A project called llama-modchip has developed a method to run the large 284 billion parameter DeepSeek-V4-Flash model on a standard 64GB CPU system without requiring any GPUs. This is achieved through a custom tensor typ…
-
llama.cpp b12203 enhances GGML CPU backend performance
The llama.cpp project has released version b12203, introducing significant performance improvements to its GGML CPU backend. This update refactors the default computation for GGML_TYPE_F16 to utilize F32 for matrix mult…
-
Strata engine enables local Qwen3.8-Flash-Next LLM deployment on consumer hardware
The Strata engine, an open-source project built on llama.cpp and ggml, enables running large language models like Qwen3.8-Flash-Next locally on consumer hardware. This engine divides models across GPU, system RAM, and S…
-
Ripple-65B Preview LLM offers 2M context window with hybrid attention
A new coding-focused LLM, Ripple-65B Preview, boasts a two-million-token context window, achieved through a hybrid attention mechanism. This architecture utilizes sliding attention for most layers, allowing for efficien…
-
llama.cpp enhances WebGPU support with new operations
The llama.cpp project has released updates focusing on WebGPU support, with recent commits adding functionality for COUNT_EQUAL and MEAN operations. These updates, assisted by Codex, aim to enhance the performance and c…
-
Stable Diffusion image generation runs on laptop iGPU
A user has successfully implemented image and video generation using Stable Diffusion on a laptop's integrated GPU, specifically an AMD Ryzen 4750U with Vega 7 graphics. Through custom patches for `stable-diffusion.cpp`…
-
VIDRAFT releases 35B LLM for CPU-only inference
VIDRAFT has released POCKET-35B-GGUF, a 35-billion-parameter large language model designed for CPU-only inference. This model is quantized and packaged in the GGUF format, making it accessible for local deployment on st…
-
llama.cpp project releases multiple pre-release updates with performance improvements
The llama.cpp project has released several pre-release versions, including b12121, b12119, b12118, b12116, b12114, b12110, b12109, b12102, b12100, and b12109. These updates address various issues and introduce new featu…
-
Developer keeps chatbot under 9-sec serverless timeout with MiniLM-Chat-6B
A developer details how they successfully deployed a chatbot on their personal website using Cloudflare Workers, adhering to a strict 9-second serverless execution limit. The solution involves using a compact MiniLM-Cha…
-
llama.cpp adds batched tensor API to ggml for performance boost
The llama.cpp project has released version b12002, introducing a generic batched tensor copy API to its ggml library. This new API is designed to improve performance by allowing multiple tensor operations to be processe…
-
llama.cpp adds OCP FP8 E4M3 quantization support
The llama.cpp project has released version b12000, introducing support for OCP FP8 E4M3 quantization. This update includes numerous fixes and improvements related to FP8 conversion, tensor handling, and backend compatib…
-
LLM KV Cache Sizing Guide for Local Models
This article explains how to accurately size the KV cache for local LLMs, which is crucial for preventing out-of-memory errors and performance degradation. It details the formula for calculating KV cache size based on m…
-
llama.cpp releases b11923 with GGML library updates
The llama.cpp project has released version b11923, which includes updates to its GGML library. These changes relax scale operations to support non-contiguous views across various backends, including CPU, CUDA, and Metal…
-
llama.cpp v0.8.0 adds LiquidAI LFM3, DeepSeek V4.1 Flash, and new audio APIs
The llama.cpp project has released version 0.8.0, introducing support for a wide array of new models including the LiquidAI LFM3 family, DeepSeek V4.1 Flash, and Nemotron-3(.5)-Embed. This update also brings significant…
-
llama.cpp releases b11857 with GGML library updates
The llama.cpp project has released version b11857, introducing several updates to its GGML library. Key changes include allowing the MUL_MAT_ID index to skip computation, refinements to Vulkan and ggml-webgpu code, and …
-
llama.cpp updates Vulkan backend with GGML precision support
The llama.cpp project has released an update, b11854, which enhances its Vulkan backend with support for GGML source precision. This update includes improvements to accumulator precision selection, handling of fp32 over…
-
b11838
A new initiative, the Global Authentic Memory Initiative (GAMI), has received €400,000 in funding from Google.org and the German Federal Government to protect historical documents from AI-generated fakes. GAMI is develo…
-
Ollama v0.40.7-rc1 optimizes MLX backend with deduplicated CUDA payloads
Ollama has released version v0.40.7-rc1, which includes optimizations for its MLX backend. This update focuses on deduplicating CUDA runtime payloads by moving shared CUDA libraries into a common payload. The changes al…
-
llama.cpp b11720 adds non-contiguous SUM support across CPU, CUDA, Vulkan, WebGPU
The llama.cpp project has released version b11720, which includes support for non-contiguous SUM operations across multiple backends. This update enhances performance by enabling these operations on central processing u…