llama.cpp ships llama.app for one-line local model installs

llama.app is a new landing and download page for llama.cpp, the open-source project for running AI models on a user's own machine. The pitch is privacy and control: the site promises no API keys, no telemetry and no limits, with users owning both their models and their conversation data. Installation is a single command, curl -LsSf https://llama.app/install.sh | sh. Run frontier AI entirely on your machine.

The page also documents a path to pairing llama.cpp with a local coding agent called Pi. After running llama serve to start a model, a user installs the pi-llama plugin with pi install git:github.com/huggingface/pi-llama, then launches Pi, which the site says will automatically discover the local model with no config and no API keys, keeping files and requests on the machine.

On hardware, the site claims llama.cpp is optimized to run on whatever is available, from a laptop to a cluster, using the same binary, the same models and the same hand-tuned kernels for every GPU and CPU.

The page lists four models a user can run through llama.cpp. Qwen 3.6, from Alibaba, is described as a next-gen natively multimodal reasoning model with dense and MoE variants that the site says rival models many times their size on coding and vision tasks. Gemma 4, from Google, is built from Gemini 3 technology and is billed as Google's most capable open model, supporting multimodal reasoning, agentic workflows and 140+ languages. GPT-OSS is described as OpenAI's first open-weight models since GPT-2, built for reasoning, agentic tasks and developer use with function calling and tool use. Gemma 3, also from Google, is built from Gemini technology, supports 140+ languages plus vision and text tasks, and offers up to 128K context for deployments ranging from edge devices to the cloud.

Key facts

llama.app installs llama.cpp locally with one command, curl -LsSf https://llama.app/install.sh | sh, and the site advertises no API keys, no telemetry and no limits.
A new pi-llama plugin connects llama.cpp to the Pi coding agent, which the site says auto-discovers the local model with no configuration once llama serve is running.
The project claims to run the same binary and same hand-tuned kernels across hardware from a laptop to a cluster, on any GPU or CPU.
Four models are listed as ready to run: Alibaba's Qwen 3.6, Google's Gemma 4 (built from Gemini 3), OpenAI's GPT-OSS (its first open-weight models since GPT-2), and Google's Gemma 3, which supports up to 128K context and 140+ languages.

Why it matters

The page's central pitch is control over data and infrastructure: no API keys, no telemetry, no limits, and the user owns both the models and the conversation data. That is the standard argument for running AI locally instead of through a hosted API, and this release packages it as a one-line install rather than a manual build.

Who it affects

Developers who want to run AI models on their own hardware without sending requests to a cloud API, and specifically those who use or want to use a local coding agent: the pi-llama plugin ties llama.cpp directly into the Pi agent so that a locally served model becomes the agent's backend with no separate configuration.

How to use it

Install with curl -LsSf https://llama.app/install.sh | sh, then run llama serve to start serving a model. For agent use, install the plugin with pi install git:github.com/huggingface/pi-llama and launch Pi, which is said to automatically discover the local model. Four models are offered out of the box: Qwen 3.6, Gemma 4, GPT-OSS and Gemma 3, covering both compact and larger deployments from a laptop to a cluster.

How solid is it

The entire text is the project's own site copy, not an independent report, so the claims are self-descriptions rather than verified results. The statement that Qwen 3.6's variants rival models many times their size on coding and vision tasks carries no benchmark numbers or named comparisons. The page gives no release date or version history for llama.app or the four listed models, no pricing or licensing detail beyond the general label open-source, and no author or maintainer name for the site itself.

Risks and caveats

Marketing language like no limits and runs on whatever you have is not backed by any hardware minimums or specifications in the source, so actual performance on a given machine is unstated. The model comparison claims (rivaling much larger models) are the vendor's characterization, not a cited benchmark, and should be read as such until independently tested.

“No API keys, no telemetry, no limits.”

— llama.app site copy