Hugging Face

    Hugging Face: find open models, check them and run them at your place.

    The model that sorts your supplier mail inside your own network comes from the Hub with a licence and a model card. We select it, mirror it to you and run it there, with Ollama or in Microsoft Foundry.

    For IT teams who want to run an open model with a checked licence in their own network or fine-tune it on their own data.

    • Hugging Face Hub
    • Enterprise Hub
    • Safetensors
    • Ollama
    • vLLM
    • Microsoft Foundry
    The botanical garden on the Brissago islands, a gardener repots labelled seedlings, a list of models on the laptop, the robot agent with a watering can beside her
    01

    Where the Hub comes in

    Muster AG wants supplier mail in three languages routed to the right department automatically, and the mail must not leave the building. IT searches the Hub for a model with a suitable licence, reads the model card, downloads the weights once through the proxy and starts them on the server in its own network. Hugging Face fits when you want to run, fine-tune or move an open model yourself and when the origin of every model has to be traceable. The same model runs locally with Ollama, on a GPU machine with vLLM or from the Hugging Face collection in Microsoft Foundry, and the application only changes its base address.

    02

    What we build with Hugging Face

    The Hub holds models, datasets and model cards with licence and origin; its libraries download, check and run them. We use it for these tasks:

    • Model selection with evidenceFor your task we shortlist three to five candidates from the Hub, check licence, origin and file format and measure them on your test set.Result: a model register with version, licence and score per model.
    • Mirror in your own networkA service account downloads the approved models once through your proxy; after that, servers and machines work from the local cache with Hub access switched off.Result: no download at runtime, every version in-house.
    • Operation on three routesLocally with Ollama straight from a GGUF repository, on GPU servers with vLLM or llama.cpp, or from the Hugging Face collection in Microsoft Foundry when the load needs an Azure subscription.Result: one application that changes where it runs by changing its base address.
    • Your own models and datasetsFine-tunes and training data live in a private organisation on the Hub, with sign-in through Microsoft Entra ID, EU storage and an audit log.Result: the team shares models internally without making them public.
    03

    Licence, access, mirror

    A test set of typical tasks and sample data. The licence of every model, read for your use, because Apache 2.0, MIT and individual vendors' licences allow different things. A service account with a token for access, because gated models pass username and email to the publisher and nobody should use a personal account for that. An approved route for downloads through your proxy and a location for the cache. And one person who owns the model register, updates and operation.

    04

    What the measurement decides

    We measure quality, response time and resource use on the test set, per candidate and per place of operation, and compare with a model in Microsoft Foundry. In the log we check that no connection to the Hub exists at runtime, that every file matches the checksum on the model card and that only models without executable code sit in the register. The decision records which model runs where and who maintains it.

    What you receive from us

    The model comparison on your tasks.

    Before you choose a model, we test the candidates on your own tasks and record which one fits what. Shown here as a sample with invented results and neutrally named models.

    Test report

    Sample

    Model comparison on three tasks

    Customer
    Muster AG, Bern (fictional)
    Tasks
    Summarising contracts, questions on internal rules, code review
    Tested
    14 to 20 October 2026
    Test data
    Anonymised documents and a test repository
    • 4models compared
    • 3tasks
    • 120test cases
    • 2models recommended

    Summary

    We tested two cloud models and two open models on three Muster AG tasks, with 120 test cases from anonymised documents. For contracts and code we recommend a cloud model with governed data processing. For questions on internal rules, an open model on your own hardware answered just as accurately and stays inside your network.

    CasesTask and resultRecommendation
    40Summarising contracts: cloud model A found every notice period, open model C missed two in appendices.Cloud model A
    50Questions on internal rules: all four answered with a source, open model C as accurately as the cloud models.Open model C, in your network
    30Code review: cloud models A and B found the planted bugs, the open models about half of them.Cloud model A or B

    To settle before use

    • Legal reviews the data processing agreement for the cloud model.
    • The open model needs a server with a GPU; IT has the quote.

    Next steps

    With your decision we set up the chosen models and rerun the test cases after every model change. The real report names each model and version.

    Samples with invented details for a fictional Muster AG.

    Sound familiar? Tell us where you stand today.

    Contact

    What's new.

    As of: October 2026

    • As of October 2026

      Any GGUF model from the Hub runs directly in Ollama.

      With ollama run hf.co/{organisation}/{repository}:{quantisation}, Ollama starts a model straight from the Hub without a Modelfile of its own; without a tag it picks Q4_K_M when present. Private repositories need the Ollama SSH key in the Hugging Face account.

      Source: Hugging Face Docs: Use Ollama with any GGUF model
    • As of October 2026

      The Hugging Face collection in Microsoft Foundry runs without a connection to the Hub.

      Managed compute in Microsoft Foundry, currently in preview, serves models from the collection out of a Microsoft registry: Safetensors only, no executable code, weights pre-staged in Azure storage, so no outbound access to the Hub. Billed per accelerator and hour, for now as a global deployment only and without an SLA.

      Source: Microsoft Learn: Managed compute in Microsoft Foundry
    • As of October 2026

      EU storage for models and datasets on the Team and Enterprise plans.

      Organisations on a Team or Enterprise plan choose whether new repositories live in the US or in the EU; without a plan they always live in the US, and Switzerland is not a region of its own. Added to that: SSO through SAML 2.0 or OIDC with Microsoft Entra ID, an audit log with export through the API and resource groups for access.

      Source: Hugging Face Docs: Storage Regions
    • As of October 2026

      Offline mode and telemetry switch in the libraries.

      With HF_HUB_OFFLINE=1 the libraries read from the local cache only and no longer contact the Hub; HF_HOME sets the storage location, HF_TOKEN the token of a service account and HF_HUB_DISABLE_TELEMETRY=1 switches off the libraries' usage data.

      Source: Hugging Face Docs: Environment variables

    What IT teams ask before the first download.

    May we use any model from the Hub commercially?

    The licence of each model decides. Apache 2.0 and MIT allow commercial use, individual vendors' licences set conditions such as a user count or a purpose, and some models are available on request only. We read the licence for your case and record the result in the model register.

    What does Hugging Face see of us?

    On download, the account that downloads, and for gated models the username and email that go to the model's publisher. In operation on your server, nothing: prompts and answers stay in-house, and with Hub access switched off and no telemetry, no request goes out at runtime either. That is why a service account downloads, not an employee.

    How do we protect ourselves against malicious model files?

    We admit only models in Safetensors format to the register, reject repositories that want to run their own code on load, and check the checksums against the model card. The Hub scans repositories for malware and pickle files, and the collection in Microsoft Foundry admits only screened Safetensors models. The download goes through your proxy and is logged there.

    Do we need the Enterprise plan?

    Not for downloading public models. Once your team stores its own fine-tunes or datasets, an organisation on a Team or Enterprise plan pays off: sign-in through Microsoft Entra ID, EU storage, audit log, resource groups and token management. Whoever uploads nothing needs only a service account with a token.

    Which runtime do we operate?

    For a server or a workstation, Ollama, which starts GGUF models straight from the Hub. For GPU servers with many simultaneous requests, vLLM, which Hugging Face itself recommends now that Text Generation Inference is in maintenance only. For an application in your Azure subscription, the Hugging Face collection in Microsoft Foundry, with Entra ID, your logs and hourly billing.

    Can we fine-tune a model on our data?

    Yes, with the transformers and peft libraries on a GPU machine at your place or in your Azure subscription; the result lives as an adapter or a model in your private organisation. Whether a fine-tune beats a better prompt with examples is what the test set shows beforehand, and often the smaller model with good instructions is enough.

    Agree on a useful next step

    Tell us about your current workflow, the systems involved and the result you need. In our first conversation, we identify what else we need to know to plan a useful next step.