Overview
PMetal is a native machine-learning platform for Apple Silicon Macs, covering model inference, long-context serving, fine-tuning, pretraining, quantization, merging, evaluation, and distributed execution. Its Rust runtime combines an MLX bridge with custom Metal and Apple Neural Engine paths tuned across M1 through M5 systems.
The platform exposes shared job specifications and event streams through a command-line interface, terminal workspace, desktop application, MCP tools, Rust builders, and Python bindings. Models can be loaded from local safetensors or the Hugging Face Hub with automatic architecture detection.
Training and Serving
Training options include supervised fine-tuning, LoRA, QLoRA, DoRA, reinforcement-learning methods, distillation, embedding training, and full-parameter pretraining. Pretraining supports tokenized binary shards, dense and mixture-of-experts models, and checkpoint resumption with optimizer state.
Serving provides continuous batching, concurrent decode slots, shared prefix caching, and OpenAI-compatible endpoints with Anthropic message support. TurboQuant uses rotation, quantization, residual techniques, and direct decode attention to compress long-context KV cache by a stated four to six times.
Model Operations
- Merges models with methods including SLERP, TIES, DARE, Fisher, and RegMean.
- Exports GGUF across multiple quantization formats and supports MLX target-bits-per-weight quantization.
- Covers major language-model families including Llama, Qwen, DeepSeek, Mistral, Gemma, Phi, Cohere, Granite, NemotronH, and GPT-OSS.
- Includes modules for BERT-family encoders, Flux diffusion, vision and speech models, CLIP, Whisper, and T5.
Apple Silicon and Distributed Workloads
Hardware-specific execution includes custom Metal kernels and M5-oriented Metal 4 and MPP dispatch for operations such as matrix multiplication, attention, mixture-of-experts workloads, distillation, LoRA, and cross-entropy. Persisted tuning allows the backend to adapt to the available hardware tier.
For multi-Mac workloads, PMetal provides local discovery, Thunderbolt-first fabric selection, ring all-reduce, pipeline harnesses, and gradient compression. The synchronized GUI, TUI, and live job logs expose training, serving, benchmarking, tokenization, and device status from the same operational layer.
Installation Options
PMetal requires macOS on Apple Silicon. It is distributed as signed prebuilt binaries, a Rust package, or source code; HTTP serving and MCP tools are optional features when installing through the Rust package ecosystem. The desktop GUI has a separate Tauri and Svelte build path.