Run AI locally.
No compromises.

vMLX is the most complete MLX inference engine for Mac. Unlike LM Studio or Ollama, vMLX provides a 5-layer caching stack (prefix + paged KV + q4/q8 quantization + continuous batching + disk cache), VL model support with full caching, speculative decoding, Mamba/SSM hybrids, 50+ auto-detected architectures, 14 tool call parsers, and built-in agentic coding tools with MCP integration. Free, no cloud connection, works on any M1+ Mac.

At 100K token context, vMLX achieves 154,121 prompt tokens/sec (cold) vs LM Studio's 686 tok/s. vMLX uses paged multi-context KV caching (concurrent conversations stay cached), while LM Studio uses single-slot caching that evicts on switch. vMLX supports up to 256 concurrent sequences vs 1 for LM Studio. All three offer OpenAI-compatible APIs, but only vMLX exposes all 23 inference parameters.

Yes. vMLX supports any MLX-compatible model from HuggingFace including DeepSeek V3, Llama 3/4, Qwen 2.5/3, Gemma 3, Mistral, Phi, and hundreds more. We also publish our own abliterated and REAP-optimized MLX models at huggingface.co/dealignai — including Qwen 3.5 VL CRACK (uncensored vision models) and Qwen 3.5 397B REAP (pruned MoE) in 4-bit and 8-bit. Models run entirely on your Mac's Apple Silicon GPU. A 16GB Mac handles up to ~20B parameters, while 64GB+ handles 70B+ models.

Prefix caching stores computed KV states from previous prompt processing. When you send a new message that shares the same system prompt or history, cached tokens are reused instantly. In benchmarks, this reduces TTFT by up to 9.7x on 2.5K context. Critical for multi-turn conversations and agentic workflows. Combined with KV cache quantization (q4/q8), you can cache even longer contexts in less memory.

No. vMLX runs entirely on your Mac with zero cloud dependency. No API keys, no subscriptions, no rate limits. Your conversations and model weights stay 100% local and private. Internet is only needed to download models initially.

Any Mac with Apple Silicon (M1, M2, M3, M4, M5 or later). More unified memory = larger models: 8GB handles ~3-7B, 16GB up to ~20B, 32-64GB handles 30-70B, and 128-512GB runs the largest open models at full precision.

Download vMLX, pick a model like Llama 3, Qwen 3, or DeepSeek V3, and use the built-in chat interface. Unlike ChatGPT, everything runs locally on your Mac — no subscription, no usage limits, no data sent to any server. You get the same chat experience with complete privacy and zero cost.

Agentic AI lets language models call external tools autonomously. vMLX has native MCP (Model Context Protocol) support with built-in tools that let your model read, write, and edit files, execute shell commands, run browser automation, search the web, and perform multi-step coding tasks — all running locally on your Mac. Configure tool iterations, tool-choice modes, and working directories. Combined with OpenAI-compatible function calling, vMLX is a complete local agentic AI platform.

Yes. vMLX exposes an OpenAI-compatible API at localhost:8000. Point any tool that supports custom OpenAI endpoints — Cursor, Continue, Aider, Open Interpreter, LangChain, or custom scripts — to your local vMLX server. All inference stays on your machine with zero latency and no API costs.

vMLX is purpose-built for Apple Silicon using the MLX framework, while Ollama uses llama.cpp. vMLX provides a 5-layer caching stack (prefix + paged KV + q4/q8 quantization + continuous batching + disk cache), speculative decoding, Mamba/SSM support, 50+ auto-detected architectures, 14 tool call parsers, and a native macOS GUI — features Ollama lacks. For Mac-native performance and developer experience, vMLX is the superior choice.

Yes. Every assistant message has a built-in text-to-speech button. Click it to listen to any response hands-free. No external services or API keys required — it uses your Mac's native speech synthesis.

Yes. vMLX supports multimodal models like Qwen VL and LLaVA. Paste or drag-and-drop images directly into the chat. Images are displayed inline with click-to-zoom, and the model can analyze and respond to visual content. All processing stays on your Mac.

Reasoning blocks show the model's chain-of-thought in a collapsible section, separate from the final answer. Supported by thinking models like DeepSeek R1, Qwen 3, and GLM-4.7 Flash. vMLX auto-detects reasoning-capable models and configures the right parser automatically — no manual setup needed.

vMLX is the only MLX engine that combines vision-language model support with a full 5-layer caching stack (prefix cache, paged KV cache, KV cache quantization, continuous batching, persistent disk cache). It also supports speculative decoding, Mamba/SSM hybrid architectures, auto-detects 50+ model architectures, has 14 tool call parsers and 4 reasoning parsers, reports cached_tokens in the OpenAI-compatible API, and uses storage-boundary quantization (full precision during generation, compressed only in storage). No other MLX app — LM Studio, Ollama, mlx-community tools, or any other — offers even a fraction of these engine capabilities.

Yes. vMLX has first-class Mamba and SSM hybrid support with a dedicated BatchMambaCache that handles batch filtering, merging, and KV quantization safety across Mamba layers. This means Mamba-based models work with continuous batching and the full caching stack. No other MLX inference app supports SSM architectures with batched inference.

Speculative decoding uses a small, fast draft model to propose candidate tokens, which the larger target model then verifies in parallel. This can significantly speed up generation without sacrificing output quality. In vMLX, configure any MLX model as the draft (e.g., a 2B model for a 30B+ target), set the number of draft tokens (default 3), and get faster responses. No other local MLX app supports this.

vMLX checks GitHub for new releases on startup and shows a dismissible banner if an update is available. Click the download link to get the latest version. There are no forced updates, no background downloads, and no telemetry — you control when and whether to update.

Yes. vMLX lets you configure a separate embedding model in the session settings. The /v1/embeddings endpoint uses this dedicated model, so you can generate embeddings for RAG pipelines or semantic search without stopping your chat model. No other local MLX app supports this.

vMLX includes 20+ built-in tools across 7 categories: File I/O (read, write, edit, copy, move, delete, list), Code Search (grep, glob), Shell (execute commands), Web Search (DuckDuckGo and Brave), URL Fetch, Git (status, diff, log, show), and Utilities (clipboard read/write, current date/time). All tools run locally with a configurable working directory and iteration limits.