Tag
The article introduces a fine-tuned Automatic1111 fork for Apple Silicon that adds Metal optimizations (e.g., Metal Flash Attention) to speed up Stable Diffusion 1.5 generation, cutting times from 8-10s to 3-7s on an M3 Pro and from 13-20s to 8-10s on an M1 Mac Mini.
Antirez's h3.c is a native Minimal inference engine for MiniMax-H3 on Apple Silicon, providing a fast, end-to-end prompt-to-video/audio pipeline with Metal optimizations and an interactive session. It is currently focused on performance and memory optimization for M3 Max and M5 Max.
Redis creator antirez releases h3.c, a pure C + Metal inference engine for MiniMax H3 that runs natively on Apple Silicon, enabling text-to-video/audio without Python, PyTorch, or ComfyUI.
Mference, a Swift + Metal inference engine, now supports Inkling-Small 276B-A12B, running it at ~2.9 tok/s on under 10GB memory, enabling large MoE models on consumer Apple hardware.
Shitty is a fast, open-source terminal emulator that uses Vulkan/Metal for GPU rendering and claims higher throughput than Alacritty, Kitty, and Ghostty in benchmarks.
TurboFieldfare is an open-source Swift+Metal runtime that runs the Gemma 4 26B-A4B model on Apple Silicon Macs using only ~2GB of RAM by streaming experts from SSD, enabling inference on 8GB machines.
The author describes implementing TurboQuant KV-cache compression into Apple's mlx-swift-lm, achieving 2.7x compression with quality on par with 8-bit, and 3-4x decode speed improvements via a fused Metal kernel.
A developer shares their implementation of a bindless GPU abstraction layer called Loon GPU, layered on Vulkan 1.3 and Metal 4, inspired by Sebastian Aaltonen's 'No Graphics API' blog post. The library uses GPU pointers, vertex pulling, and bindless texture heaps to simplify modern graphics APIs.
QuixiAI released QuixiCore, a family of native high-performance AI kernel libraries for modern accelerators, with standalone implementations for CUDA, Metal, ROCm, XPU, and Gaudi backends, all sharing a common contract but no shared code.
QuixiAI rebrands ThunderKittens and ThunderMittens into QuixiCore-CUDA and QuixiCore-Metal, creating a unified family of cross-platform kernels for AI workloads.
QuixiAI ported ThunderKittens to Metal, enabling kernel support on MPS and MLX for training models on Mac.
The author describes building a GPU-based display backend for Emacs using Metal on macOS and OpenGL on Linux, improving rendering performance and enabling new effects like video playback and animated cursors, without modifying the core redisplay engine.
Ported the EXL3 LLM codec to run on Apple Silicon via Metal, achieving high prefill and generation speeds on M5 Max (e.g., ~600 tok/s prefill, 17-80 tok/s gen on various models).
zml/llmd now runs fully on Apple's Metal API, serving 8 simultaneous requests at full bf16 precision, with continuous batching and other modern features.
Rigel is an empirical characterization of Apple's Metal 4.1 tensor compute path on the M4 Max GPU, revealing that fp8 matmul2d is emulated (not accelerated), the operation executes entirely on GPU shader cores without a dedicated matrix datapath, and reconstructing the opaque cooperative tensor fragment layout.
safetensors 0.8.0 release brings major performance improvements: direct copy into Metal MTLBuffers with dlpack for 2-3x faster loading and OOM fix on macOS, plus GIL-free serialization for faster multi-file saves.
GenBench is a free iOS app that lets users download, run, and benchmark GGUF models on iPhone/iPad using llama.cpp and Metal, with features like offline chat, standardized benchmarks, and a global leaderboard.
User shares their experience using llama.cpp with the GGUF Q4_K_M quantized version of Gemma-4-12b on a Mac, achieving local inference speed of about 36 tok/s and memory usage of about 10GB.
An interactive map visualizing the subgenres of heavy metal music.
Releases a kernel on Hugging Face that accelerates MaxSim late-interaction retrieval by using tiled scoring with SIMD group matrix operations (Metal and WMMA), achieving 3–5× speedup over the naive implementation.