Local AI models, on your own hardware.
Even LLMs runs open models on your machine through our own llama.cpp engine, GPU-accelerated. It recommends what will run well, downloads it safely, and makes it usable by you and your agents. No cloud, no token bills, nothing leaves the device.
Recommended for your machine
Even reads your RAM, CPU, GPU, and VRAM, then tells you which models will run, at which quantization and context, with an expected tokens per second. No guessing, no out-of-memory surprises.
Search any model
Browse a curated catalog or search Hugging Face for any GGUF model. See every quantization with its size and whether it fits your machine, then download. Each file is checksum-verified.
GPU-accelerated
The engine uses Metal on Mac and CUDA or Vulkan elsewhere, with a CPU fallback. Even picks the right build for your hardware and the optimal flags, so models run as fast as the machine allows.
Agents delegate to it
A paid agent like Claude Code can hand a subtask to a local model: bulk classification, JSON extraction, first drafts, or private and offline work, then use the result. Cheap when it can be, frontier when it counts.
A terminal it can drive
Open a terminal whose driving agent is a local model, with Even's tools attached. A fully local agent loop for work you want kept on the machine.
Private and offline
Everything runs on-device. No prompts or code leave your machine, no per-token billing, and it works with no internet once a model is downloaded.
Detect, recommend, run, in a few clicks.
Even owns the whole path so you do not assemble a runtime by hand.
Pick a model and press run.
Even installs the right engine on first use, downloads the model, starts it with the best flags for your hardware, and exposes a local endpoint your agents and terminals can use.
› qwen2.5-coder-7b · fits · ~62 tok/s · GPU
› llama-3.1-8b · fits · ~58 tok/s · GPU
› llama-3.3-70b · needs ~48 GB
✓ running on 127.0.0.1 · agents can delegate