Open-Source AI Models: When They Beat Proprietary and When They Do Not

Agentic Business Design

15 December 2025 | By Ashley Marshall

The gap between open-source and proprietary AI models has collapsed. Gemma 4, Qwen 3.6, and Llama 4 now match or exceed closed models on many enterprise benchmarks - at a fraction of the cost. The question is no longer whether open-source works. It is when it makes sense for your business.

The Open-Source Landscape in April 2026

Twelve months ago, choosing an open-source model for production work felt like a compromise. Today, six major open-weight model families compete at the frontier: Gemma 4 from Google, Qwen 3.6 Plus from Alibaba, Llama 4 from Meta, Mistral Small 4, OpenAI's gpt-oss-120b, and GLM-5 from Zhipu.

Five of these six use mixture-of-experts (MoE) architecture, meaning active parameter counts range from just 5.1 billion to 40 billion - enabling single-GPU inference for models with total parameters running into the hundreds of billions. The practical impact: you can run genuinely capable models on a single NVIDIA A100 or even an RTX 4090.

Crucially, the licensing picture has cleared up. Apache 2.0 and MIT licences now cover the majority of leading open models, including Gemma 4, Qwen 3.6, Mistral Small 4, gpt-oss, and GLM-5. That eliminates the legal grey areas that previously made enterprise adoption risky.

Where Open-Source Wins Clearly

Data sovereignty and privacy. If your business handles sensitive customer data, financial records, or health information, running models on your own infrastructure means nothing leaves your network. For UK firms navigating GDPR and the upcoming Data Protection and Digital Information Act, this is not just convenient - it is often the simplest compliance path.

Cost at scale. Proprietary API calls (GPT-5, Claude 4, Gemini Ultra) typically cost between $5 and $60 per million tokens. Self-hosted open models on leased GPU infrastructure run at roughly $0.10 to $1.50 per million tokens once you factor in compute costs. For businesses processing millions of documents, support tickets, or internal queries, that difference compounds fast.

Customisation and fine-tuning. Open-weight models can be fine-tuned on your proprietary data. A legal firm can train Llama 4 on decades of case law. A manufacturing company can specialise Mistral on equipment maintenance logs. Proprietary APIs offer limited fine-tuning options and your training data goes through their infrastructure.

Latency control. Self-hosted models eliminate network round trips to external APIs. For real-time applications - chatbots, code completion, in-app suggestions - this matters.

Where Proprietary Models Still Lead

Complex multi-step reasoning. For tasks that require chaining together multiple reasoning steps - financial analysis combining market data with regulatory constraints, or multi-document legal review - the top proprietary models (GPT-5, Claude Opus 4) still outperform open alternatives. The gap is narrowing, but it exists.

Multimodal breadth. While Gemma 4 and Qwen 3.6 handle text and images well, proprietary models generally offer more polished vision, audio, and video capabilities in a single unified model. If your workflow involves analysing mixed media, proprietary APIs are currently smoother.

Zero infrastructure overhead. Not every business wants to manage GPU servers. Proprietary APIs are a single HTTP call - no provisioning, no patching, no capacity planning. For small teams or early-stage projects, the operational simplicity is genuine value.

Rapid iteration at the frontier. Anthropic, OpenAI, and Google ship model updates frequently. If your competitive advantage depends on always having the newest capabilities, proprietary APIs deliver that without migration work on your end.

The Hybrid Approach Most UK Businesses Should Take

The smartest enterprises are not choosing one or the other. They run open models for internal, high-volume, data-sensitive workloads and reserve proprietary API calls for external-facing, high-stakes tasks where absolute quality matters most.

A practical split might look like this:

• Internal knowledge base and document search: Self-hosted Llama 4 or Qwen 3.6 with RAG
• Customer-facing chatbot: Proprietary API (GPT-5 or Claude) with fallback to open model
• Code generation and review: Gemma 4 or gpt-oss locally, with proprietary API for complex refactors
• Regulatory document analysis: Fine-tuned Mistral on-premises for data sovereignty

This hybrid pattern keeps costs manageable, maintains data control, and ensures quality where it matters most.

How to Evaluate Which Path Suits Your Business

Before committing to either direction, work through these questions:

1. What is your monthly inference volume? Below 10 million tokens per month, proprietary APIs are probably cheaper than maintaining infrastructure. Above that, the economics shift towards self-hosting.
2. Does your data leave your network? If regulatory or client requirements prohibit external API calls, open-source is your only practical option.
3. Do you have ML engineering capacity? Running open models well requires someone who understands GPU provisioning, model serving frameworks (vLLM, TGI), and monitoring. If you do not have that skill set, proprietary APIs remove that burden.
4. How specialised is your domain? Generic tasks (summarisation, translation, general Q&A) work brilliantly with open models out of the box. Highly specialised domains benefit from fine-tuning, which requires open weights.

There is no universal right answer. The right answer depends on your data sensitivity, volume, team capability, and budget.

Frequently Asked Questions

Are open-source AI models really free to use commercially?

Most leading open-weight models in 2026 use Apache 2.0 or MIT licences, which permit commercial use without royalties. However, you still pay for the compute infrastructure to run them. Always verify the specific licence for any model you deploy.

Can a small UK business realistically self-host an AI model?

Yes, but it depends on the model size. Smaller models like Gemma 4 9B run on a single consumer GPU. Larger models need cloud GPU instances from providers like AWS, Azure, or Lambda Labs. Managed deployment platforms like Replicate and Together AI also offer pay-per-use hosting of open models.

How do open-source models handle GDPR compliance?

Self-hosting open models means data never leaves your infrastructure, which simplifies GDPR compliance significantly. You control data processing, retention, and deletion entirely. With proprietary APIs, you rely on the provider's data processing agreements and trust their compliance claims.

What is mixture-of-experts and why does it matter?

Mixture-of-experts (MoE) is an architecture where only a fraction of the model's total parameters activate for each query. A 200 billion parameter MoE model might only use 20 billion parameters per inference, dramatically reducing compute costs and enabling deployment on smaller hardware.