A year ago, the conversation about open-source AI was dominated by a simple narrative: Meta’s Llama models were good but behind OpenAI and Anthropic. Open-source was the scrappy underdog. Proprietary was the premium choice.

That narrative is now outdated. In March 2026, the competitive landscape between open-source and proprietary AI models is far more nuanced, and the implications for businesses choosing between them are different than most people assume.

Where Open Source Stands Today

Let’s look at the facts rather than the marketing.

Meta’s Llama 4 models, released in early 2026, represent a significant leap. The largest variant performs competitively with GPT-4 class models on most standard benchmarks. For many practical applications — text generation, summarisation, code generation, reasoning — the quality gap between Llama 4 and the leading proprietary models is negligible.

Mistral has established itself as the European open-source powerhouse. Their models consistently punch above their weight in efficiency — achieving strong performance with fewer parameters, which translates directly to lower compute costs.

Alibaba’s Qwen series and several Chinese open-source models have improved dramatically, particularly for multilingual applications. For businesses operating across Asia-Pacific, these models offer capabilities that English-centric models can’t match.

The Hugging Face Open LLM Leaderboard tracks performance across standardised benchmarks, and the trend is unmistakable: the gap is shrinking on almost every metric.

The Real Competitive Dynamics

Here’s where the simple “open vs. closed” narrative breaks down. The competition isn’t just about model quality. It’s about at least five different dimensions, and open-source wins on some while proprietary wins on others.

Raw model capability: Proprietary models still hold an edge at the absolute frontier — the hardest reasoning tasks, the most nuanced instruction following, the most complex multi-step problems. But that edge has narrowed from a gulf to a gap. For 80% of business use cases, open-source models are good enough.

Cost: Open source wins decisively here. Running a Llama 4 model on your own infrastructure or through a cloud provider costs a fraction of API calls to OpenAI or Anthropic. For high-volume applications — chatbots processing millions of queries, document processing pipelines, code generation tools — the cost difference is significant.

Customisation: This is where open source wins most dramatically. When you have the model weights, you can fine-tune the model on your specific data, modify its behaviour for your specific use case, and optimise it for your specific hardware. You can’t do any of this with a proprietary API. For specialised applications — medical terminology, legal language, industry jargon — fine-tuned open-source models routinely outperform general-purpose proprietary ones.

Deployment control: Open-source models can run on-premises, in private clouds, in air-gapped environments, or on edge devices. If your data can’t leave your infrastructure — and for many Australian enterprises, particularly in health, defence, and government, it can’t — open source is your only option.

Safety and alignment: Proprietary models currently have an edge in safety guardrails and alignment. The major providers invest heavily in testing, red-teaming, and alignment research. Open-source models can be fine-tuned to remove safety guardrails, which is both a feature (for legitimate research) and a concern (for misuse). The AI Safety Institute in Australia has flagged this as an area requiring policy attention.

What Businesses Should Actually Consider

If you’re an Australian business evaluating AI options, the open-source vs. proprietary question should be framed around your specific constraints, not around abstract benchmarks.

Choose proprietary if: You need the absolute best model performance and can afford API costs. You want managed safety and compliance features. You don’t need to fine-tune. You’re comfortable with data leaving your infrastructure.

Choose open source if: Cost matters at scale. You need to fine-tune for your specific domain. Your data can’t leave your infrastructure. You want to avoid vendor lock-in. You have (or can hire) the technical capability to deploy and manage models.

Choose both if: You’re serious. Many sophisticated organisations use proprietary APIs for prototyping and customer-facing applications where quality matters most, while running open-source models for internal tools, batch processing, and specialised applications where cost and customisation matter more than marginal quality improvements.

The Maturity and Strategy Questions

The enterprise open-source stack has matured substantially. vLLM for model serving, LangChain for orchestration, and RAG frameworks have moved from experimental to production-ready. Cloud providers now offer managed open-source model hosting alongside proprietary options.

But the most important consideration is dependency risk. If you build entirely on a proprietary API, you’re dependent on that provider’s pricing and technology decisions. Open-source models eliminate that dependency. You can switch models, run multiple simultaneously, and maintain control. In a market as fast-moving as AI, that flexibility has genuine strategic value.

The Honest Assessment

Open source hasn’t “won.” The best proprietary models are still better for the most demanding applications. But for the vast majority of practical business use cases, open-source models are now good enough — and they offer advantages in cost, customisation, and control that proprietary models can’t match.

The smart money isn’t betting on one side winning. It’s building infrastructure that can work with both, shifting workloads between open and proprietary models based on requirements rather than ideology.

That’s a more boring narrative than “David vs. Goliath.” But it’s the accurate one.