Ka1zen
A native macOS app for running open large language models entirely offline through Apple's MLX framework, with a llama.cpp/GGUF fallback engine, vision and audio input, image generation, thinking-mode visibility, and speculative decoding for faster output.
Ka1zen is a native macOS chat application that runs open LLMs such as Qwen, Gemma, DeepSeek, Mistral and Llama locally through Apple's MLX framework, with no account, cloud call or telemetry; conversations, documents and embeddings stay on the device. A second engine runs GGUF models directly through a bundled llama-server for models not yet supported by MLX, reaching feature parity with the MLX path for text, vision, tool calls and thinking output.
Vision models let users drop images into the chat, and a block-diffusion language model (Google's DiffusionGemma) runs through MLX using parallel denoising instead of left-to-right generation. A speculative-decoding mode, called Fast Mode, works on both engines with byte-identical output and covers both dense and mixture-of-experts models, including a variant that roughly doubles code-generation throughput on some models. An audio mode lets Gemma 4 hear and answer spoken questions, and a thinking mode shows the model's reasoning before its final answer.
Web search is available through Brave Search with DuckDuckGo as a fallback.