Languages
RustTypeScriptC++
Design goal
ease-of-useedge/mobilemulti-modellow-latencycross-platform
Scheduling
parallel-model-loading
Backends
cpu-avx512cpu-avx2cudarocmhipvulkanmetal
Notable features
A Tauri desktop app (Rust core, web front-end) with an OpenAI-compatible local server on localhost:1337, so other applications can use whatever Jan has loaded.Backend variants are a first-class build concept: engine builds are selected with JAN_ENGINE_VARIANT from the tokens cpu, vulkan, metal, cuda12, cuda13, and hip/rocm, composable with '-'.Ships an mlx-server component alongside the llama.cpp engine, so Apple Silicon has a native MLX path in addition to llama.cpp/Metal.Custom assistants, MCP integration for agentic use, and cloud-provider connectors (OpenAI, Anthropic, Mistral, Groq, MiniMax) alongside local models.Available as Windows exe, macOS universal dmg, Linux deb and AppImage, plus Microsoft Store and Flathub.
Hardware caveats
The build-time engine variant is a real fork in the road: picking JAN_ENGINE_VARIANT=cpu gives you a CPU-only build and no GPU path at all. There is no runtime backend switching comparable to llama.cpp's --gpu flag across variants.Windows build requires VS 2022 Build Tools (MSVC x64 workload + Windows SDK), LLVM, Ninja and CMake on PATH; the bundled llama.cpp engine builds with Ninja + clang-cl. The CUDA Toolkit is needed only for the cuda12/cuda13 variants.Windows-specific build trap: an nvcc 'Could not open output file' error during the llama.cpp build means the build path crossed Windows' 260-character MAX_PATH limit, because nvcc does not honour the long-path opt-in. The build script relocates the engine build tree to %LOCALAPPDATA%\jan-engine automatically; if it still happens set JAN_ENGINE_BUILD_DIR to a short path or move the checkout closer to the drive root.make dev must be run from Git Bash on Windows, since make dispatches recipes through sh — a plain cmd.exe will not work.macOS 13.6+ required; the published minimum RAM guidance is 8GB for 3B models, 16GB for 7B and 32GB for 13B. Apple Silicon builds also need the MetalToolchain Xcode component.Windows 10+ with GPU support for NVIDIA/AMD/Intel Arc is stated as the minimum for a decent experience, but the actual backend you get depends entirely on which engine variant was built.
Best for
A desktop ChatGPT replacement for a non-technical user who wants a GUI and a working download-and-go experience.Apple Silicon users who want an app that uses MLX rather than only llama.cpp.Offline, private use where cloud connectors are optional rather than the default.Feeding a local OpenAI-compatible endpoint to other apps at localhost:1337.Users who want assistants and MCP in one place instead of assembling separate tools.
Avoid for
Headless/server deployments — this is a desktop app with a sidecar server, not a serving stack.Anyone needing to change backends at runtime; the backend is chosen at build time.Benchmarking or performance work — Jan adds a desktop layer, not kernel speed.Multi-user concurrency.Linux Arm64, which is documented only as a how-to issue link rather than a first-class release artifact.
notes
The engine-variant set (cpu, vulkan, metal, cuda12, cuda13, hip/rocm) is the most accurate description of Jan's real backend matrix, so supported_backends reflects that rather than a claim that all are present in any one build. GitHub reports the license as NOASSERTION while the README states Apache 2.0; recorded as Apache-2.0 per the project's own statement, with the discrepancy noted. Jan is a UI layer over llama.cpp plus MLX; the kernel-level backend detail belongs to the llama.cpp and MLX records. ||| MAINTENANCE (dated fact, GitHub API, 2026-10-04): last commit 2026-10-02 (0.1 months before this record), latest release v0.8.4 published 2026-07-23, 184 contributors (distinct committers including anonymous), open_issues_count 542 (issues PLUS pull requests - see [[ev-github-open-issues-count-includes-pull-requests]]), not archived, 44796 stars. See [[ev-maintenance-status-is-four-dated-facts]] for how to read this fact set. ||| DEPLOYMENT: NO EVIDENCE OF PRODUCTION DEPLOYMENT FOUND as of 2026-10-04 - repositories, READMEs, docs and issue trackers were searched. See [[ev-production-evidence-is-scarce-and-nearly-always-single-site]] for the grading scheme; note that for an artifact-evaluation or paper-code release, non-deployment is documented intent rather than a negative finding.