Hugging Face's Top Collaborations and Model SetupPublic

Hugging Face has built its position as the central hub of the AI ecosystem largely through strategic collaborations. Here are the most impactful ones:

BigScience / BLOOM — This is probably the most historically significant collaboration in terms of what it represented. Launched in 2021, Hugging Face coordinated over 1,000 researchers from 66 countries and 250 organizations to build BLOOM, a 176-billion-parameter multilingual LLM trained on 46 languages. It was one of the first major demonstrations that open, community-driven research could produce frontier-scale models, and it set the template for how HF would position itself as a neutral platform for open AI.

AWS — The February 2023 strategic partnership with Amazon was a turning point commercially. AWS became HF's preferred cloud provider, integrating deeply with SageMaker, Trainium, and Inferentia. AWS is the most popular place to run Hugging Face models, and the partnership has grown exponentially since launch. It also signaled that HF was serious about enterprise revenue, not just community.

Microsoft / Azure — Microsoft deepened its collaboration with HF to make over 10,000 Hugging Face models available to Azure developers, with integration into Azure AI Foundry. This was notable given Microsoft's simultaneous deep partnership with OpenAI — it showed that even Microsoft recognized the need to support the open ecosystem alongside its proprietary bets.

NVIDIA — NVIDIA and HF partnered to connect the HF community with NVIDIA's AI computing platform, including DGX Cloud. HF launched "Training Cluster as a Service" powered by NVIDIA infrastructure. This gave the community access to serious compute, which had been a key bottleneck for open-source model development.

Google — Beyond Google participating in HF's $235M Series D round (at a $4.5B valuation), Google Cloud integrated with HF for training and deploying models on GKE and Vertex AI. Hundreds of thousands of HF users are active on Google Cloud monthly.

Apple / Core ML — Apple has published models and datasets directly on Hugging Face, including Core ML optimized models for on-device inference. This collaboration reinforced HF as the default distribution channel even for a company as closed as Apple historically has been.

Meta (indirect but massive) — While not a formal partnership in the traditional sense, Meta's decision to distribute Llama models through Hugging Face has been enormously consequential for both parties. Llama became one of the most downloaded model families on the Hub and cemented HF as the place where major model releases happen.

The $235M Series D round itself, with participation from Google, NVIDIA, Amazon, and Salesforce, was arguably a collaboration signal in its own right — it meant all major cloud and AI infrastructure players had a stake in HF's success and neutrality.

The common thread across all of these: HF positioned itself as the Switzerland of AI, a neutral platform everyone could use, which made it rational for competitors (AWS, Google, Azure) to all partner with them simultaneously.

Sources:

• Hugging Face - Wikipedia
• Hugging Face and AWS partner to make AI more accessible
• Microsoft and Hugging Face deepen generative AI partnership
• NVIDIA and Hugging Face to Connect Millions of Developers
• Hugging Face and Google partner for open AI collaboration
• Apple embraces open-source AI with Core ML models on Hugging Face
• Hugging Face raises $235M from investors
• BigScience and Hugging Face's AI Language

Based on the screenshot, the command to launch the model server is:

llama-server -hf ggml-org/gemma-4-26b-a4b-it-GGUF:Q4_K_M

And then to onboard it with OpenClaw:

openclaw onboard --non-interactive \
  --auth-choice custom-api-key \
  --custom-base-url "http://127.0.0.1:8080/v1" \
  --custom-model-id "ggml-org-gemma-4-26b-a4b-gguf" \
  --custom-api-key "llama.cpp" \
  --secret-input-mode plaintext \
  --custom-compatibility openai \
  --accept-risk

The first command starts a local llama.cpp server serving the Gemma 4 26B model (quantized to Q4_K_M), and the second registers it with OpenClaw so you can interact with it through that interface. You'd run the llama-server command first, wait for it to be ready, then run the openclaw onboard command in a separate terminal.