Runtimes
Five programs that run a model on your own machine. This table answers what their own websites do not: whether they send usage data without asking, who else they call, which address they listen on — and, above all, whether anyone outside the company can check any of it. Everything here was read in the published source code and the official documentation on 25 August 2026; nothing comes from watching network traffic.
The four questions that matter
| Runtime | Telemetry out of the box | Other calls home | Which address it listens on | Can it be audited? |
|---|---|---|---|---|
| Ollama | No third-party analytics, but it does call home verified by us | Update check every hour (cannot be turned off) and remote catalogue every 4 h (OLLAMA_NO_CLOUD) verified by us | 127.0.0.1:11434, no key, permissive CORS verified by us | Yes — MIT, public code verified by us |
| LM Studio | States it has none vendor declared | Update check (no documented switch) and model search vendor declared | 127.0.0.1:1234, no key, but CORS off and API keys available vendor declared | No — closed source, reverse engineering forbidden verified by us |
| llama.cpp | None at all verified by us | None: it does not check for updates or anything else verified by us | 127.0.0.1:8080, no key, CORS reflects any origin verified by us | Yes — MIT, public code verified by us |
| MLX (mlx-lm) | None of its own; inherits the download library's verified by us | Model downloads only (HF_HUB_DISABLE_TELEMETRY / HF_HUB_OFFLINE) verified by us | 127.0.0.1:8080, no key, CORS wide open verified by us | Yes — MIT, public code verified by us |
| vLLM | Yes, on by default — heartbeat every 10 min to stats.vllm.ai verified by us | Model download identifying itself to Hugging Face as vLLM; local metrics endpoint verified by us | 0.0.0.0:8000 — every interface, no key verified by us | Yes — Apache-2.0, public code verified by us |
Every cell in these four columns says how we know it: ● verified by us in the primary source, ◐ vendor declared and not independently verified, ○ we could not verify it. Read the last column first: one of these five cannot be audited by anyone outside the company, and that turns its whole row from fact into promise. None of this comes from watching network traffic — we read the published source code and the official documentation.
Everything else
| Runtime | What it solves | Where it runs | What it accelerates | Formats | Steps to the first token | Licence | What it keeps on your disk |
|---|---|---|---|---|---|---|---|
| Ollama | A model answering in two minutes, no parameters to touch. | Windows · macOS Apple Silicon · macOS Intel (CPU only) · Linux (no desktop app) | CUDA, ROCm, Metal, Vulkan (on by default), CPU. No NPU. | GGUF and safetensors | 2 steps, no compiling | MIT | Yes: chats in an unencrypted SQLite database |
| LM Studio | Everything on screen, no command line anywhere. | Windows (AVX2) · macOS Apple Silicon · not macOS Intel · Linux (AppImage only) | CUDA, ROCm, Metal, Vulkan, CPU — from a January 2025 source. NPU: not verified. | GGUF and MLX | 3 steps, no terminal | Proprietary. Free at work; may not be resold as a service | Yes: conversations as plain JSON |
| llama.cpp | Total control and the widest hardware support, terminal included. | Windows · macOS Apple Silicon · macOS Intel (CPU only) · Linux · Android, iOS, RISC-V, s390x | CUDA (binary on Windows only), ROCm, Metal, Vulkan, SYCL, OpenVINO, OpenCL, CPU | GGUF only | 2 steps, no compiling in the normal case | MIT | Not on the server: in your browser (IndexedDB) |
| MLX (mlx-lm) | Maximum native performance on an Apple Silicon Mac, Python included. | macOS Apple Silicon · Linux for the framework · Windows undocumented · not macOS Intel | Metal (unified memory), CUDA on Linux, CPU. Neural Engine: no. | MLX and safetensors. Not GGUF | 2 steps, pip | MIT | Conversations are not saved |
| vLLM | Serving a model to a hundred people, not to yourself. | Linux · macOS Apple Silicon (experimental, CPU, compiling) · Windows only through WSL · macOS Intel: not verified | CUDA (first class), ROCm, CPU. Metal only through an external plugin. | safetensors and quantisations; GGUF only through an external plugin | 2 steps, pip — but no interface at all | Apache-2.0 | Not saved (it has no chat interface) |
These columns come from the same primary sources. Each figure carries its own provenance inside the runtime’s own sheet, quote and link included.