What is Open-Weight AI? A clear guide

What are model weights?

Every AI model — whether GPT-4, Claude, or Llama — is built on a neural network. During training, the network adjusts billions of numerical values called "weights" to learn patterns from its training data. These weights are the model's knowledge: they determine how it processes input, generates output, and responds to prompts.

Think of weights as the model's "brain cells." A closed model like GPT-4 keeps its weights completely private — you can only interact with it through an API, and you never see what's inside. An open-weight model publishes its weights publicly, typically on platforms like Hugging Face, so anyone can download them, inspect them, modify them, or run them on their own hardware.

The size of a model's weights correlates with its capability: a 7B (billion parameter) model is generally less capable than a 70B model, which is less capable than a 350B+ model. But capability also depends on training data quality, architecture, and training methodology — size alone isn't everything.

Open-weight vs. open-source vs. closed

These terms are often used interchangeably, but they mean different things:

Closed models

Models like OpenAI's GPT series and Anthropic's Claude are completely closed. You don't have access to the weights, the training data, or the architecture. You interact exclusively through an API, accept the provider's terms of service, and pay per token. You have no visibility into how the model makes decisions, no ability to audit it, and no way to deploy it independently.

Open-weight models

Models like GLM-5.2, Qwen, and Llama publish their weights publicly. Anyone can download them and run them locally. However, "open-weight" doesn't always mean "open-source" in the traditional sense — the license may restrict commercial use, require attribution, or limit deployment scale. For example, Llama's license (Llama 3.3 Community License) allows commercial use but has conditions on the number of users. Most Qwen3-generation models use the permissive Apache-2.0 license — but not all: Qwen2.5-VL and the Qwen3.8 Max and Flash models ship under Qwen's own licenses, so check each model page.

Truly open-source models

A small subset of open-weight models are also fully open-source: the training pipeline, the dataset details, and the architecture are all documented and reproducible. Models like Ai2's OLMo, which gives access to its training data, code and weights, fall into this category. Most models marketed as "open-source" are actually "open-weight" — the weights are available but other aspects remain proprietary.

This distinction matters for compliance. An open-weight model with a restrictive license may not be suitable for enterprise deployment, even if the weights are publicly available. Always check the license before production use.

Why open-weight matters for enterprise

Enterprise teams are increasingly choosing open-weight models over closed alternatives for several concrete reasons:

1. Vendor lock-in avoidance

With a closed model, you're locked into one provider's API, pricing, and roadmap. If they raise prices, change terms, or discontinue a model, you have no fallback. Open-weight models let you switch infrastructure providers without changing your model — and many can be self-hosted as a final fallback.

2. Data privacy and compliance

When you send data to a closed model API, it passes through the provider's infrastructure. For closed models, that typically means a US provider, subject to the CLOUD Act wherever its servers are. With open-weight models served by EU-headquartered providers — like OVHcloud or Scaleway — inference stays outside the CLOUD Act's direct reach. This is critical for GDPR compliance and increasingly required by enterprise contracts.

3. Independent verification

With open-weight models, you can verify a model's capabilities through independent benchmarks, run your own tests, and inspect the weights for bias or security concerns. You're not relying on the provider's claims about performance or safety — you can measure it yourself.

4. Cost predictability

Open-weight models are served by multiple providers, creating price competition. You can choose between managed APIs with transparent pricing (like Frontière AI's prepaid billing) and self-hosting for high-volume workloads. Read our analysis of self-hosting vs. API cost for the math.

5. Customization

Open-weight models can be fine-tuned on your domain-specific data, adapted with custom system prompts, or modified for specialized tasks. This level of control is impossible with closed models.

The most capable open-weight models in 2026

The open-weight landscape has matured significantly. Here are notable open-weight models across sizes, all available on Frontière AI:

These models are all served on EU-sovereign providers through Frontière AI's catalog, with transparent pricing; our benchmarks page publishes the six-criteria verification status of each live model, each criterion ticked only once its check has been run, recorded and passed. See also our articles on the best open-source AI models in 2026, the best open-source reasoning models, and open-source models with tool calling.

How to deploy open-weight models

There are three main deployment paths for open-weight models, from simplest to most complex:

1. Managed API (simplest)

Services like Frontière AI give you an OpenAI-compatible endpoint to access open-weight models without managing any infrastructure. You get one API key, transparent pricing, and — in Frontière AI's case — inference by EU-controlled providers on the models labeled EU sovereign. This is the fastest path to production and the most practical for teams without ML infrastructure expertise.

2. Self-hosted gateway

Tools like LiteLLM (routing proxy), vLLM (inference engine), or Ollama (local development) let you serve open-weight models on your own infrastructure. See our guide to open-source AI gateways for a detailed comparison. This path gives you maximum control but requires GPU infrastructure and operational expertise.

3. Cloud provider hosting

Cloud providers like AWS, Google Cloud, and Azure offer managed inference for popular open-weight models. This middle ground gives you managed infrastructure while keeping the model weights in your control — but the cloud provider's own jurisdiction and terms still apply. For EU sovereignty, OVHcloud AI Endpoints and Scaleway Managed Inference are the relevant options.

Limitations of open-weight AI

Open-weight models have closed the gap with closed models in many areas, but limitations remain:

• Closed models still lead in certain domains. GPT-4o and Claude 4 still outperform open-weight models on some benchmarks, particularly in coding, complex reasoning, and long-context understanding. The gap is narrowing — Qwen3.5 397B and GLM-5.2 are competitive on many tasks — but it's not zero.
• Hardware requirements. Running a 300B+ model locally requires enterprise-grade GPUs (multiple A100s or H100s). Smaller models (7B-30B) can run on consumer hardware with quantization, but they sacrifice capability.
• Licensing complexity. Not all open-weight models have permissive licenses. Some restrict commercial use, require revenue thresholds, or limit the number of users. Always verify the license before production deployment.
• Support and documentation. Closed models come with dedicated support, SLAs, and extensive documentation. Open-weight models rely on community support (though model labs like Qwen and GLM are investing more in documentation).

Frontière AI — managed access to open-weight models

Frontière AI bridges the gap between the flexibility of open-weight models and the simplicity of a managed API. You get access to capable open-weight models — GLM-5.2, Qwen3.5 397B, Qwen3 235B, DeepSeek V4 Flash, and more — through a single OpenAI-compatible endpoint. Every model is labeled by jurisdiction: the EU-sovereign ones run at EU-controlled providers (OVHcloud, Scaleway), and the fast-access ones at US companies.

Unlike self-hosting, you don't manage GPUs or model downloads. Unlike closed model APIs, inference on sovereign models stays with providers you can name and that sit under European jurisdiction; the gateway itself runs on Vercel's Paris region, a US company, in transit only — see our privacy policy. With prepaid billing starting from €10, you spend only what you top up: calls are refused once the balance reaches zero (a call already running can take it slightly below, and the difference comes off your next top-up).

See our guide to AI API aggregators for a comparison with other managed options, and our deep dive on GDPR vs. the CLOUD Act for the compliance context.

FAQ