Why Open Source AI Models Are Winning (Even When Proprietary Models Are Better)
On most standard benchmarks — MMLU, HumanEval, GSM8K — the best proprietary models from OpenAI, Anthropic and Google consistently outperform the best open-weight models from Meta, Mistral and the open source community. GPT-4 scores higher on the bar exam than Llama 3. Claude 3.5 writes more reliable code than Mistral Large. Yet the open source AI ecosystem is not just surviving — it is thriving, attracting billions in venture capital, powering thousands of commercial applications and building a community of developers that proprietary models cannot match. The reason is not benchmark performance. It is control, cost and customisation.
The Economic Argument for Open Models
Running a proprietary model like GPT-4 through an API costs approximately $30-60 per million input tokens and $60-120 per million output tokens for the most capable tier. For a customer support chatbot handling 10,000 conversations per day, that translates to roughly $500-1,500 per day in API costs — or $180,000-550,000 annually. Running an equivalent open-weight model like Llama 3 70B on dedicated GPU infrastructure requires an upfront investment of roughly $30,000-80,000 for the hardware, plus ongoing electricity and maintenance costs of perhaps $5,000-10,000 per year. After the first year, the open model is dramatically cheaper. For scale-ups and enterprises with predictable inference workloads, the economics favour self-hosting open models by an order of magnitude.
The cost advantage extends to fine-tuning. With a proprietary API, fine-tuning is limited to whatever the provider offers — typically a few epochs of supervised fine-tuning on a narrow prompt-completion format at a significant markup. With an open model, developers can fine-tune on custom data, use parameter-efficient methods like LoRA and QLoRA that reduce training costs by 90% or more, and deploy multiple specialised fine-tunes for different tasks. A startup called Nous Research (notable for the Hermes series of fine-tuned models) has demonstrated that community fine-tunes of open models can outperform much larger proprietary models on specific tasks. The ability to specialise is a structural advantage that proprietary APIs cannot replicate.
The Control Argument
When a company builds its product on GPT-4, it is entirely dependent on OpenAI for pricing, availability, model behaviour and content moderation policies. When OpenAI changes its API — deprecating a model version, modifying response behaviour, adjusting pricing — every dependent application must adapt. The November 2023 chaos at OpenAI, when the board briefly fired CEO Sam Altman and most employees threatened to quit, was a stark reminder of the concentration risk in proprietary AI. If OpenAI had imploded that week, thousands of startups would have lost access to their core infrastructure overnight.
Open models eliminate this dependency. A company that deploys Llama 3 on its own infrastructure controls every aspect of the model’s behaviour. It can modify the system prompt, adjust the temperature, add retrieval-augmented generation (RAG) for domain-specific knowledge and deploy the model behind its own firewall. For regulated industries — healthcare, finance, defence — where data sovereignty and auditability are non-negotiable, open models are often the only viable option. The US Department of Defense has funded multiple projects to deploy open-weight AI models in classified environments where cloud API calls are impossible. The European Union’s AI Act imposes strict requirements on high-risk AI systems, and the ability to audit and explain model behaviour is far easier with open models than with proprietary black boxes.
The Innovation Argument
The pace of innovation in the open source AI ecosystem is extraordinary. When Meta released Llama 2 in 2023 under a commercially permissive licence, the community immediately began experimenting: quantising the model to run on consumer hardware, fine-tuning it for specific domains, combining it with other models through mixture-of-experts approaches, and building an entire ecosystem of tools around it — from inference servers (vLLM, llama.cpp) to fine-tuning frameworks (Axolotl, Unsloth) to serving infrastructure (Together AI, Fireworks). None of this ecosystem development would be possible around a proprietary API. The collective intelligence of thousands of open source developers, iterating at internet speed, consistently finds optimisations and use cases that the original model creators never anticipated.
This dynamic creates a flywheel: more developers building on open models attracts more investment in open model infrastructure, which reduces costs, which attracts more developers. Mistral, the French AI company, has explicitly built its business around the intersection of open and proprietary — releasing capable open-weight models that serve as a gateway to its premium API services. Meta, whose core business is social media and advertising, not cloud AI, has invested billions in open-weight models (Llama) as a strategic move to shape the ecosystem and reduce dependence on competitors Google and Microsoft. The motivation is not altruism — it is competitive strategy. The result, however, is a thriving open AI ecosystem that is structurally more resilient and more innovative than any single proprietary provider could be. Open source AI is winning not because it is better — but because it is yours.
