What it is
llamafile turns an open model into one file you can download and run. It combines llama.cpp with Cosmopolitan Libc, so the same file runs on macOS, Windows, Linux, FreeBSD, OpenBSD and NetBSD on both x86-64 and ARM64 CPUs.
It started as a Mozilla Builders project and is maintained by Mozilla.ai. After a gap between 0.9.3 in May 2025 and 0.10.0 in March 2026, the 0.10 series moved to a new build system that tracks upstream llama.cpp closely; the docs say it brings newer model support but may lack some 0.9 features and has not been tested on every GPU and platform.
It also includes whisperfile, a single-file speech-to-text tool based on whisper.cpp. The project is Apache 2.0; its changes to llama.cpp and whisper.cpp are MIT.
Our review
llamafile solves distribution, not inference. One executable holds the engine and, optionally, the weights; on macOS and Linux you mark it executable and run it, on Windows you rename it to .exe. It has about 26,500 GitHub stars and has shipped seven 0.10 releases since March 2026, the latest on 15 September.
The portability has edges. Windows will not run an executable over 4 GB, so anything beyond a small quantised model needs the bare binary plus a separate GGUF file, or WSL. x86-64 CPUs must have AVX. GPU support covers Metal, CUDA, ROCm and Vulkan, but the docs call the AMD and Windows paths best-effort, and if a backend fails to load llamafile quietly falls back to the CPU unless you force one with --gpu.
The security design is more careful than most local runners. In --server mode it applies a pledge sandbox on Linux and OpenBSD that blocks outbound connections and file writes, and --confine-reads hides everything outside the model's directory. The docs are candid about the gaps: there is no sandbox at all with a GPU enabled, none on macOS or Windows, and none in the default combined mode. The server listens on port 8080, and --host 0.0.0.0 opens it to the network.
For daily work, llama.cpp itself gets new models and flags first, and Ollama handles model management better. llamafile's niche is the hand-off: a workshop USB stick, an offline field laptop, a demo for a colleague who will never install Python. In that niche the 4 GB Windows ceiling is the constraint that matters most.
Where it earns its place
A developer-relations engineer running a 40-seat workshop
Get every attendee chatting with a local model in five minutes, whatever laptop they bring
Hand out a small pre-built llamafile such as Qwen3.5 0.8B under 4 GB so it runs on Windows too. No Python, no package manager; the cost is choosing a model small enough for the weakest laptop.
A field researcher working without internet
Transcribe interviews and summarise notes on an offline laptop
Carry a whisperfile and a llamafile on the drive and run both on CPU. Nothing leaves the machine; expect slow output on older CPUs, and AVX is required on x86-64.
An IT lead at a 50-person firm piloting local AI under a no-cloud policy
Give staff an internal chat endpoint without sending data to a vendor
Run --server on a Linux box with --confine-reads, keep weights in their own directory and front port 8080 with a reverse proxy. The sandbox only applies in CPU mode, so a GPU server needs its own isolation.
Before you commit
- Setup
- Download a .llamafile, chmod +x and run it (rename to .exe on Windows); or download the bare binary and point -m at any GGUF. The 0.8B sample model answers within minutes on a laptop CPU.
- Integrations
- llama.cpp server and web UI with an OpenAI-compatible endpoint on port 8080; --cli and --chat terminal modes; optional built-in tools and an MCP proxy; whisperfile for transcription.
- Data and privacy
- Runs locally; no telemetry mentioned in the docs. Server mode on Linux blocks outbound connections and writes, but GPU mode, macOS, Windows and the default combined mode run unsandboxed, and a llamafile from an untrusted source can strip the sandbox.
- Cost at scale
- Free, with no paid edition from Mozilla.ai. Hardware is whatever runs the model: an AVX x86-64 or ARMv8 CPU with enough RAM for the weights, optionally a Metal, CUDA, ROCm or Vulkan GPU.
- Limits
- 216 open issues and pull requests; 4 GB Windows executable cap; bundled CUDA lacks IQ-quant kernels; multi-GPU may break on AMD; 0.9 examples untested on 0.10.
What's good
- One file, no install
- Runs on six operating systems, x86-64 and ARM64
- Server sandbox blocks outbound network on Linux
- Includes speech-to-text via whisperfile
What's not
- 4 GB executable limit on Windows
- No sandbox with GPU enabled, or on macOS and Windows
- Silent fallback to CPU if a GPU backend fails
- 0.10 rebuild dropped some 0.9 features
Best for
People who need to share or run a local model with zero setup, including offline.
Alternatives
llamafile: common questions
- Is llamafile worth it?
- The simplest way to hand someone a working local model, but less capable than the llama.cpp it wraps. We rate it 3.9 out of 5.
- How much does llamafile cost?
- Free and open source under the Apache 2.0 licence, with MIT-licensed changes to llama.cpp.
- What are the downsides of llamafile?
- 4 GB executable limit on Windows; no sandbox with GPU enabled, or on macOS and Windows; silent fallback to CPU if a GPU backend fails; 0.10 rebuild dropped some 0.9 features.
- Who is llamafile best for?
- People who need to share or run a local model with zero setup, including offline.
- How much does llamafile cost for a team or in production?
- Free, with no paid edition from Mozilla.ai. Hardware is whatever runs the model: an AVX x86-64 or ARMv8 CPU with enough RAM for the weights, optionally a Metal, CUDA, ROCm or Vulkan GPU.
- Is llamafile safe for company data?
- Runs locally; no telemetry mentioned in the docs. Server mode on Linux blocks outbound connections and writes, but GPU mode, macOS, Windows and the default combined mode run unsandboxed, and a llamafile from an untrusted source can strip the sandbox.
- What are the best llamafile alternatives?
- The closest alternatives we have reviewed are llama.cpp and Ollama.
Answers checked Oct 23, 2026.