mlx

mlx-vlm v0.6.4 is released with support for 5 new model families, TTS/STT endpoints, and significant performance improvements including TurboQuant and continuous batching.

The author used Codex to optimize DeepSeek V4 Flash 8-bit MLX on oMLX, achieving approximately 1.6x prefill and 3x decode speedup.

OpenMed privacy-filter v2 using nemotron and MLX 8-bit achieves 755 tokens per second on a Mac, redacting 1,152 PII identifiers across 22 categories from a 13,000-token clinical file without data leaving the machine.

Prism ML releases Ternary-Bonsai-27B-mlx-2bit, a ternary-quantized 27B-parameter language model that achieves ~95% of FP16 performance while fitting in ~7.2 GB, enabling full reasoning on laptops.

A free, fully local voice cloning and TTS tool powered by Qwen3-TTS and Kokoro runs on Apple Silicon via MLX, enabling studio-quality audiobook generation from PDFs without cloud services.

Ollama announces that Gemma 4 is now nearly 90% faster on Apple Silicon using MLX, thanks to improved multi-token prediction enabled by default, with automatic tuning to avoid slowdown.

QuixiAI ported ThunderKittens to Metal, enabling kernel support on MPS and MLX for training models on Mac.

A detailed guide on fine-tuning a 7B instruction model using LoRA with MLX on Apple Silicon, with example outputs showing stylistic adaptation to high-fantasy literary registers.

A guide on building a large language model from scratch using Apple's MLX framework.

A collection of 650+ Apache-2.0 licensed biomedical NER and de-identification models that run on-device via MLX, achieving 30-40x faster inference than PyTorch-CPU on an M3 Max with identical outputs.

Rapid-MLX is a local LLM inference tool optimized for Apple M-series chips. Built on the MLX framework, it achieves 2 to 4 times faster inference than Ollama, supports multiple models, tool calling, and an OpenAI API-compatible interface.

GLM 5.2, an open-weight AI model comparable to top closed models, has been released and is now running on MLX on two Mac Studios (M3 Ultra).

Announces the release of a Config-I quantization of MiniMax-M3 on MLX, using 2-bit experts and 4-bit attention to reduce the 427B MoE model from 869GB to ~167GB, though the quant is untested and requires a patch for mlx_lm.

The react-native-executorch library now integrates Google's Gemma 4 model, enabling fully offline, GPU-accelerated inference in React Native apps using Vulkan on Android and MLX on Apple Silicon.

MLX-LoRA-Studio is a native macOS app for fine-tuning LLMs on Apple Silicon, offering a user-friendly interface and support for various training algorithms including SFT, DPO, and QAT. It is fully open-source and allows local, private fine-tuning without cloud dependency.

oMLX, a MLX server for local AI, now supports the standard Hugging Face cache model directory, simplifying model loading.

A tweet highlights an excellent WWDC video by Angelos Kath on building local agentic AI with MLX, noting rapid progress in open-weight models and hardware capabilities.

MTPLX V1 is a native Mac app that bundles the MTP speculative decoding engine for MLX models, offering features like model conversion via Forge, built-in chat, benchmarking, and support for smaller models. It achieves over 2x speedup with mathematical exactness.

Yagil Bubrovnik presented at WWDC, demoing LM Studio's upcoming clustering feature on stage, crediting the MLX team for their work.

Cohere officially launches North Mini Code, a coding model, with weights available on Hugging Face and deployment support for vLLM and MLX.