> For the complete documentation index, see [llms.txt](https://docs.nvcl.ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.nvcl.ai/get-started/requirements.md). # Requirements What a machine needs to run NVCL AI: Linux on x86-64 with a supported NVIDIA GPU, or a Mac with Apple Silicon. Nothing else needs to be installed beforehand: no CUDA toolkit, no Python, no Docker; for the NVIDIA driver, see [Operating system and driver](#operating-system-and-driver). `nvcl-ai check` checks this machine, and `nvcl-ai models` shows which models and formats fit. {% hint style="info" %} NVCL AI is in beta and free for personal use, with a free NVCL AI account: see [Get NVCL AI](/overview/request-access.md). NVCL AI, the `nvcl-ai` command and these docs are updated all the time, so commands, menus and requirements may change. {% endhint %} ## GPU, memory and disk NVCL AI supports NVIDIA H100, H200, B200, B300 and RTX PRO 6000 GPUs (see [Hardware](/reference/hardware.md)); on another GPU, `nvcl-ai install` says it has no engine for it, or that no model fits, and stops. What each model and format needs: | Model | Format | GPUs | Host RAM | Download | | ------------------------------------------------- | ------------------------------------------------------------------- | ------------------------------------------------ | ----------------------------------- | ----------------------------------- | | Qwen3.8-27B | Standard (default) | One or two, each with 40 GB or more | | 35 GB | | Qwen3.8-27B | Compact, Throughput | One or two H100 or H200 | | 34 GB (Compact), 35 GB (Throughput) | | Qwen3.8-27B | Exact | One or two H100 or H200, each with 64 GB or more | | 65 GB | | Qwen3.8-Flash-Next | Standard (its only format, a build of its own) | One H200 (one GPU with 135 GB or more) | 96 GB | 161 GB | | DeepSeek-V4.1-Flash | Exact (the published weights) | Two H200 | 280 GB | 510 GB | | GLM-5.3 | Compact | Two H200 | 330 GB, and a CPU with AVX-512 BF16 | 544 GB | | GLM-5.3 | Compact, a build of its own for two B200 (`nvcl-ai` 0.5.5 or newer) | Two B200, each with 176 GB or more | 230 GB, and a CPU with AVX-512 BF16 | 523 GB | | MiMo-V2.6-Flash-RL (MiMo V2.6 Flash in `nvcl-ai`) | Exact | Two H200 | | 342 GB | * For Qwen3.8-27B, the prefix cache (earlier conversations kept in host RAM for reuse) takes up to a quarter of host RAM; for Qwen3.8-Flash-Next, 32 GB. * Two GPUs must be of the same architecture, and each needs the full amount in the table: two GPUs do not halve it. * `nvcl-ai` counts sizes in GiB, so the 34-35 GB downloads show as 32-33 GB and the 65 GB one as 60 GB. * Parts the formats share are downloaded once. The engine is a further download of up to 2.3 GB (for GLM-5.3, 32 MB on H200 and 35 MB on B200). `nvcl-ai` checks the free space before downloading, counting only what it still has to download ([Model files you already have](/get-started/install.md#model-files-you-already-have)); when it is short, `nvcl-ai install` names a folder on a roomier disk, and `--models-dir DIR` puts the model files there. ## Mac NVCL AI runs on Macs with Apple Silicon (M1 or newer) and macOS 13.3 or newer; `nvcl-ai install` checks both before it downloads a model. A Mac needs no NVIDIA driver or CUDA. What each format needs: | Model | Format | Unified memory | Download | | ----------- | ----------------------- | -------------- | -------- | | Qwen3.8-27B | Small (the Mac default) | 32 GiB or more | 19 GB | | Qwen3.8-27B | Compact | 36 GiB or more | 23 GB | | Qwen3.8-27B | Standard | 48 GiB or more | 28 GB | * `nvcl-ai` shows these downloads as 18, 22 and 26 GB (it counts in GiB). Parts the formats share are downloaded once. The engine is a further download of about 6 MB. * What Qwen3.8-27B serves on a Mac: [Supported models](/reference/models.md). ## Operating system and driver On Linux: | Item | Requirement | | ------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | | OS | Linux on x86-64 with glibc 2.35 or newer (Ubuntu 22.04, Debian 12, RHEL 10 or newer). `nvcl-ai` checks it before it downloads anything or offers a driver: an older system, such as Ubuntu 20.04, RHEL 8 or RHEL 9, stops with exit code 1. | | NVIDIA driver | 580 or newer; `nvcl-ai check` checks it. When it is missing or too old, `nvcl-ai install` offers to install it, with sudo and only after you say yes (see [Administrator rights](/get-started/install.md#administrator-rights)). In a container the driver is the host's and is never installed. | | CUDA toolkit | Not needed: the engine brings the CUDA libraries it uses. | ## Network * `nvcl-ai` downloads the catalog, the engine and the models from our download site over HTTPS; the engine and the models need your free NVCL AI account (see [Sign in](/get-started/install.md#sign-in)). * It reaches no Python package sites or container registries. Only a driver install or an install of missing system tools, each after your yes, reaches the package repositories its commands name. * Behind a proxy, set `HTTPS_PROXY`; `NO_PROXY` is read too. * When you open a model to the network, it asks the download site for this machine's address, to print where the server can be reached. * It sends no usage data. Next: [Install](/get-started/install.md).