MLX
A machine learning framework developed by Apple and optimized for Apple Silicon chips, enabling efficient local execution of large language models on Mac hardware.
Core Facts
MLX is specifically optimized for Apple Silicon's unified memory architecture.
MLX是Apple专为Apple Silicon芯片设计的机器学习框架，能充分利用M系列芯片统一内存架构（UMA）的优势
AMD's ROCm and Intel's oneAPI still have significant gaps compared to CUDA in terms of ecosystem completeness, community support, and performance tuning tools.
Timeline (last 90 days)
oMLX is an LLM inference service built on top of MLX, supporting continuous batching, two-tier memory+SSD KV caching, and compatibility with the OpenAI and Anthropic APIs.
Qwen3.5, Qwen3.6, and Qwen3.8 gained faster prefill and speculative decoding support on MLX.
Ollama 0.19 was released on March 30 and switched to the MLX backend on Apple Silicon, with decoding speed jumping from 58 tokens/s to 112.
John Rocky's Apple Silicon LLM benchmark recorded NVIDIA Nemotron 3 Nano at 159.7 tokens/s on MLX and 86.2 tokens/s on LLAMA.cpp, but the quantization of the two runs was not consistent.
In a fair 4-bit comparison, MLX with D-Flash 2 speculative decoding ran at 16.6 tokens/s, versus Ollama Q4's 5.7 tokens/s.
Ash's LLM Picker project collected 92 public Mac measurement data points, showing that on most models MLX leads LLAMA.cpp by only 1 to 1.2 times, with the large gaps mainly concentrated in the Qwen 3.5 and 3.6 series.
Running Gemma 4 26B on an M3 Max, MLX is 65.0 tokens/s and LLAMA.cpp is 63.9 tokens/s, essentially a tie between the two.
Jordan Ilchif tested Qwen 3.8 27B on a 32GB M4 MacBook Air, with Ollama's NVFP4 build running at 17.2 tokens/s, pure MLX at 6.6 tokens/s, and LLAMA.cpp at 4.4 tokens/s, but the formats were different.
Commit #18976 enables MLX to load the Kev model, and to simultaneously load its FP32 decision head and release metadata.
MLX was open-sourced at the end of 2023, designed specifically for the unified memory architecture of M-series chips, with core features of lazy computation and a unified memory model.
40 more timeline events
All Facts (20)
MLX is specifically optimized for Apple Silicon's unified memory architecture.
85%VerifiedMLX是Apple专为Apple Silicon芯片设计的机器学习框架，能充分利用M系列芯片统一内存架构（UMA）的优势
80%VerifiedAMD's ROCm and Intel's oneAPI still have significant gaps compared to CUDA in terms of ecosystem completeness, community support, and performance tuning tools.
75%VerifiedM1's unified memory architecture allows the CPU and GPU to physically share the same memory, eliminating the need for explicit data copying and requiring only the passing of pointers or virtual addresses.
75%VerifiedApple open-sourced the MLX framework in late 2023, designed specifically for Apple Silicon.
75%VerifiedApple Silicon's Neural Engine shares a Unified Memory Architecture with the CPU and GPU, eliminating the bottleneck of data transfer between different memory pools in traditional heterogeneous computing.
75%VerifiedApple's MLX framework is a machine learning framework designed specifically for M-series chips by Apple in late 2023; it uses a unified memory architecture to eliminate memory copy overhead and can run 70-billion-parameter models on the M2 Ultra (with up to 192GB unified memory).
65%VerifiedOllama v0.40.0 enables the MLX runtime by default on Apple Silicon chips to perform model inference.
75%UnverifiedApple Silicon M-series unified memory architecture allows CPU and GPU to share the same physical memory, eliminating data copy overhead beneficial for LLM inference
90%Unverifiednpcpy深度适配Apple Silicon，在MLX架构下运行LoRA，rank拉到128，配合Flash Attention 2.0
80%UnverifiedMLX supports lazy evaluation computation graphs, triggering computation only when results are actually needed.
70%UnverifiedThe quantized MLX KV cache reduces prompt memory usage by up to 74% compared to before.
70%UnverifiedMLX is a deep learning framework open-sourced by Apple's machine learning research team in late 2023, designed specifically for Apple Silicon, employing a lazy computation strategy and natively supporting a unified memory model.
70%Partially Verified统一内存架构使CPU、GPU与NPU共享同一块内存池，避免数据在处理单元间反复搬运的开销
65%UnverifiedFor small-to-medium models that fit within 32 GB of VRAM, the RTX 5090 leads (approximately 45% on Gemma 4 12B and over 75% on GPT-OSS 20B), but the lead on Qwen 3 27B drops from 47% at Q4 to 23% at Q8.
60%UnverifiedMLX is a machine learning framework designed by Apple specifically for Apple Silicon and excels in the prefill stage; GGUF is a universal quantization format in the llama.cpp ecosystem with stronger cross-platform compatibility
60%UnverifiedoMLX is an LLM inference service built on top of MLX, supporting continuous batching, two-tier memory+SSD KV caching, and compatibility with the OpenAI and Anthropic APIs.
50%UnverifiedQwen3.5, Qwen3.6, and Qwen3.8 gained faster prefill and speculative decoding support on MLX.
50%UnverifiedJohn Rocky's Apple Silicon LLM benchmark recorded NVIDIA Nemotron 3 Nano at 159.7 tokens/s on MLX and 86.2 tokens/s on LLAMA.cpp, but the quantization of the two runs was not consistent.
50%UnverifiedRunning Gemma 4 26B on an M3 Max, MLX is 65.0 tokens/s and LLAMA.cpp is 63.9 tokens/s, essentially a tie between the two.
50%