Hugging Face & Together.ai — Model Marketplaces

The Open-Weight Model Revolution

The foundation modelFoundation ModelA large AI model trained on broad data that can be adapted to many tasks. Examples include OpenAI's GPT, Anthropic's Claude, Google's Gemini and Meta's Llama families. Banks evaluate these for capabilities, safety, and regulatory fit.See glossary landscape is not limited to proprietary models from OpenAI, Anthropic, and Google. A parallel revolution is happening in open AI, where powerful models are freely available for anyone to download, deploy, and modify. These are usually called "open-weight" models: you get the trained model itself, though usually not the data it was trained on. For banking institutions, this ecosystem offers something most proprietary models cannot: complete control over your AI infrastructure.

Two companies have become the gateways to this ecosystem. Hugging Face operates the world's largest repository of open AI models. Together AI provides the enterprise-grade infrastructure to run those models efficiently. Together, they represent a compelling alternative -- or complement -- to proprietary model vendors.

Hugging Face: The GitHub of AI

Hugging Face Hub is to AI models what GitHub is to source code: a central repository where researchers and companies publish, share, and collaborate on models. With more than 3 million public models available (August 2026), Hugging Face has become the default distribution platform for the open AI community.

What Hugging Face Offers

Model repository: Browse, evaluate, and download models ranging from compact 1-billion-parameter models to massive models with hundreds of billions of parameters. The most-used families today include Qwen (Alibaba), DeepSeek, Llama (Meta), Mistral, Gemma (Google), NVIDIA Nemotron, OpenAI's gpt-oss and Cohere's Command A+ -- plus thousands of fine-tuned variants.

Model cards: Standardized documentation for each model -- training data, performance benchmarks, intended use cases, known limitations. This transparency is valuable for banking model risk management, where you need to document model provenance and characteristics.

Transformers library: An open-source software library that provides a unified interface for loading and running models. Your data science team can switch between model architectures with minimal code changes.

Spaces: Interactive demo environments where you can test models before committing to deployment. Evaluate a model's performance on your specific use cases before investing in infrastructure.

A Change of Ownership

In September 2026 NVIDIA -- the dominant supplier of AI chips -- agreed to acquire Hugging Face, for about $11.9 billion plus up to $1 billion in retention equity. The deal is expected to close in the first half of 2027, subject to regulatory approval, and NVIDIA has committed to keeping the platform open. For banks, it is a concentration-risk point worth noting: the neutral marketplace for open models is set to belong to the company that also supplies most of the hardware they run on.

Together AI: Enterprise InferenceInferenceThe process of running a trained model to generate predictions or outputs from new input data. Inference cost, latency, and throughput are key factors in enterprise AI deployment.See glossary Infrastructure

Downloading an open-weight model is free. Running it efficiently at enterprise scale is not. Together AI, which now describes itself as the "AI Native Cloud," bridges this gap by providing optimized inference infrastructure for open models. It raised $800 million at an $8.3 billion valuation in July 2026, and in August 2026 signed a multi-year agreement with IBM to run open-model inference on IBM Cloud -- a familiar enterprise route in for many banks.

The Inference Challenge

Running large AI models requires specialized hardware -- typically NVIDIA GPUs with sufficient memory and compute power. A 70-billion-parameter model might require 4-8 high-end GPUs just to load into memory. Managing this infrastructure -- provisioning, scaling, monitoring, optimizing -- requires specialized ML engineering talent.

Together AI handles this complexity by offering:

• Managed inference endpoints: Run any major open-weight model through a simple APIAPI (Application Programming Interface)A standardized interface that allows software systems to communicate. In AI, APIs let your applications send prompts to a model and receive generated responses programmatically.See glossary call, without managing GPU infrastructure
• Fine-tuning platform: Fine-tuneFine-TuningThe process of further training a pre-trained model on a specific dataset to specialize its behavior for a particular domain or task, such as banking compliance language.See glossary open-weight models on your own data, creating banking-specific model variants
• Cost optimization: Together AI says its optimized serving infrastructure delivers lower inference costs than running models on raw cloud GPU instances -- a vendor claim to test against your own workloads
• Model variety: Switch between different open-weight model families through the same API interface

Banking Use Cases for Open-Weight Models

Open-weight models offer distinct advantages for specific banking scenarios:

On-Premises Deployment for Sensitive Data

When processing customer financial data, proprietary trading strategies, or regulatory examination materials, many banks require that no data leave their infrastructure. Open-weight models can be deployed entirely on-premises -- in your own data center, on your own hardware, with no external API calls.

Cost Control at Scale

Proprietary model APIs charge per token. For high-volume use cases -- processing millions of customer communications, classifying thousands of transactions daily, or summarizing weeks of market data -- these costs add up. Open-weight models deployed on your own infrastructure have fixed costs (hardware and operations) regardless of volume, which may be more economical at scale.

Fine-Tuning for Banking-Specific Tasks

Open-weight models can be fine-tuned on your institution's data -- regulatory filings, credit memos, compliance correspondence -- to create models that deeply understand banking terminology and conventions. This level of customization is typically not available with proprietary models.

Vendor Independence

Relying on a single proprietary model provider creates concentration risk. If that provider changes pricing, deprecates a model version, or modifies their data handling policies, your institution is affected. Open-weight models reduce this dependency -- you hold the model weights and can run them indefinitely. But the supplier of the next version can still change course, and the hubs you download from are part of your software supply chain, so the dependency shrinks rather than disappears.

Evaluating Open-Weight Models for Banking

When evaluating open-weight models, apply the same rigor you would to any model under your institution's Model Risk Management framework:

The Practical Path Forward

For a banking institution exploring open-weight AI, a practical approach is:

1. Evaluate on Hugging Face Spaces: Test promising models against your actual use cases before any infrastructure investment
2. Prototype with Together AI: Build a proof-of-concept using Together AI's managed inference -- no GPU procurement needed
3. Fine-tune for your domain: Use Together AI's fine-tuning platform to create a banking-specific model variant
4. Decide on deployment: Based on results, choose between continued managed inference (Together AI) or bringing the model in-house on your own infrastructure

This graduated approach lets your institution build confidence and capability incrementally, without committing to large infrastructure investments upfront.