Why open models matter
The case for an open AI future—and why access matters more than ever
Open-source models are faster, more customizable, more private, and pound-for-pound more capable. They are doing things with $100 and 13B params that we struggle with at $10M and 540B. And they are doing so in weeks, not months.
This quote, pulled from a leaked internal Google memo in May 2023, was meant as a warning. At the time, OpenAI’s GPT-4 had been out for three months and was dominating headlines, while Google, Meta, and everyone else scrambled to catch up.
OpenAI looked unstoppable, but in the background, something else was happening. A quiet revolution was brewing in the open-source community—one that would soon upend assumptions about who gets to build powerful AI, and how fast they can do it.
When Meta’s Llama models began circulating online, developers, hobbyists, and researchers seized the opportunity. These open models were quickly fine-tuned, optimised, shrunk, and deployed on everything from laptops to smartphones, and even Raspberry Pis. Within months, open models that anyone could download and run were achieving capabilities that rivalled (and sometimes surpassed) the most advanced closed systems.
Nearly two years later, the anonymous Google employee’s prophecy has come true: open models are not just catching up—they’re changing the game. This article explores how we got here, why it matters, and what the open future of AI might look like.
The Rise of Open Models
In early 2023, the narrative was simple: the most advanced AI would always be locked behind proprietary walls, available only through cloud APIs controlled by a handful of tech giants.
When an early version of Meta’s Llama model began circulating online, the open-source community didn’t hesitate. Developers immediately began porting it to run on local machines, customising it for niche applications, and building tools that would kickstart a wave of innovation. The model that was never supposed to be public became the catalyst for a new kind of AI development—fast, decentralised, and radically accessible.
Meta responded not with lawsuits or crackdowns, but by embracing the moment. In July 2023, it officially released Llama 2 under a more permissive license. It wasn’t yet on par with GPT-4, but it didn’t need to be. The model was good enough to be useful and open enough to evolve quickly. Within months, developers had extended its context window, fine-tuned it for specific use cases, and optimised it to run on consumer hardware. By April 2024, Llama 3 and the massive Llama 3.1 405B model had significantly narrowed the performance gap with GPT-4 and Claude 3.
And it wasn’t just Meta pushing the envelope. Chinese company DeepSeek stunned the industry with the release of its R1 model, an open-weight reasoning model that matched or outperformed OpenAI’s o1 on several benchmarks, just weeks after o1 launched.
Today, it’s no longer unusual to see open models competing head-to-head with their closed counterparts. More importantly, they’re doing it on vastly smaller budgets, with broader participation, and at much faster iteration speeds. The old idea—that only billion-dollar labs could build meaningful AI models—is falling apart.
When state-of-the-art technology becomes infrastructure
At first, AI models were the product. GPT-3 wasn’t just a component you could use—it was the experience. But as the technology matures, we’re beginning to see a shift: large language models are starting to behave less like products and more like infrastructure or components of larger systems.
Soon, nobody will care whether your app is powered by GPT-4, Claude 3, or Llama, just like nobody asks if an app is built with Node or Python. People care whether it works and whether it delivers value. AI is being abstracted away, turned into a layer of intelligence that sits quietly underneath the products people actually use.
This is exactly what happened with earlier waves of software infrastructure. Linux, once a fringe operating system, now powers the majority of servers on the internet and billions of other devices. MySQL, Postgres, NGINX—all open-source tools —quietly replaced their proprietary counterparts because they were flexible, cost-effective, and improved rapidly thanks to community contributions.
We’re watching the same playbook unfold in AI.
Need image generation? You don’t need to build a startup—you can grab Stable Diffusion and fine-tune it. Need summarisation, classification, or document extraction? There’s a Llama or Mistral variant on Hugging Face ready to go.
Even OpenAI seems to recognise this. Their product roadmap is increasingly focused on building on top of models, with tools like Sora, agents for research, and recently coding agents. But these, too, are being cloned in open-source form, often within weeks. What was once considered state-of-the-art is now just another building block. Even more, I am confident that when we reach AGI, an open AGI will be created and released on GitHub and Hugging Face within months, if not weeks.
When AI becomes infrastructure, openness wins. Flexibility, auditability, and independence become far more valuable than temporary performance edges. Just like Linux didn’t need to be better than Windows on day one—it only needed to be good enough to start the compounding loop of adoption and improvement.
Open models are crossing that threshold. And once that happens, there’s no going back.
The strategic and technical advantages of open models
While performance benchmarks often dominate headlines, real-world adoption of AI hinges on a mix of cost, control, and flexibility, and this is where open models shine.
1. Cost efficiency
Closed models come with usage-based pricing and escalating API costs. For startups or enterprises operating at scale, these costs can spiral quickly. Open models, by contrast, are often free to use. Downloading a model from Hugging Face or running it through a local interface like Ollama costs nothing beyond your own compute. For many teams, this means skipping the subscription model and external dependency entirely.
2. Total ownership and control
Open models give you something proprietary models never will: ownership. You’re not renting intelligence from a black-box API. You have the model—you can inspect it, modify it, and run it on your own infrastructure. That means no surprise deprecations, pricing changes, or usage limits.
Control also translates to trust. In regulated industries like finance, healthcare, and defence, organisations need strict control over how data flows through their systems. Closed APIs can create unacceptable risks, both in terms of data sovereignty and operational transparency.
With open models, organisations can ensure privacy by running models locally or on tightly controlled cloud environments. They can audit behaviour, integrate with their security frameworks, and version control their models like any other part of their tech stack.
3. Fine-tuning and specialisation
Open models are not one-size-fits-all, and that’s a strength. Whether it’s through full fine-tuning or lightweight adapters like LoRA, developers can adapt models to domain-specific tasks. Legal documents, biomedical data, and financial transactions—open models can be trained to understand specialised language and nuance far better than general-purpose APIs.
Even the model size can be adjusted to fit the task. DeepSeek’s R1 model, for instance, comes in distilled versions from 1.5B to 70B parameters—optimised for everything from edge devices to high-volume inference pipelines. Llama or Google’s Gemma family of open models also come in different sizes, and developers can choose which one is the best for the task.
4. Performance where it counts
Yes, top closed models may still lead in some reasoning-heavy benchmarks. But open models are closing the gap fast, and in many common workloads, they’re already at parity or ahead.
Most users aren’t asking their models to solve Olympiad-level maths problems. They want to summarise documents, structure unstructured text, generate copy, write emails, and classify data. In these high-volume, low-complexity tasks, open models perform exceptionally well, and with much lower latency and cost.
Add to this community-driven optimisations like speculative sampling, concurrent execution, and KV caching, and open models can outperform closed models not just in price, but in speed and throughput as well.
5. The rise of edge and local AI
This compute decentralisation is especially relevant for industries that need local inference: healthcare, defence, finance, manufacturing, and more. When models can run on-site or on-device, they eliminate latency, reduce cloud dependency, and strengthen data privacy.
Open models enable this shift in ways closed models never will. No API quota. No hidden usage fees. No unexpected rate limits.
The performance-per-pound advantage is compounding in open models’ favour, and enterprise users are noticing. The value is no longer just in raw capability, but in deployability.
The safety and trust argument
Critics of open-source AI argue that open models are not safe. The concern is understandable—when anyone can run and modify a powerful language model, how do you ensure it won’t be misused? But this argument overlooks a core principle of software development: transparency isn’t a liability—it’s a safeguard.
Security through visibility
In the software world, this idea is known as Linus’s Law: “Given enough eyeballs, all bugs are shallow.” In other words, when code is open, it can be inspected, audited, and improved by the community. That’s why open-source software runs the internet, powers critical infrastructure, and is trusted by governments and banks worldwide. The most widely used open-source software has been scrutinised by so many eyes that many bugs have already been found and resolved. And if a new bug is found, it is usually patched very quickly.
The same logic applies to open AI models. When weights, architectures, and training data are visible, developers can scrutinise them, identify vulnerabilities, and patch issues—all without waiting on a vendor. Open models can be tested and hardened in the open, across thousands of use cases and deployment environments. That’s not a bug. It’s a feature.
Closed models and the drift problem
Contrast this with proprietary APIs, where developers have no insight into how models work or when they change. Closed models can drift over time: prompts that work one month may yield different results the next. And because these models are black boxes, there’s no way to audit what changed or why.
For businesses building on top of LLMs, this is a problem. When the foundation of your product is unpredictable, trust erodes. And when the provider can modify that foundation without notice—or even revoke access—the risk compounds.
Control builds trust
Ultimately, trust comes from control. You can’t truly trust what you can’t see or manage. For organisations with sensitive data or strict regulatory requirements, this becomes non-negotiable. Closed APIs require trusting a third party—not only with the model, but with the data itself, and with access to how your application behaves.
With open models, the data stays inside your walls. You decide how models are deployed, what safety checks to apply, and when (or whether) to upgrade. You’re not beholden to a company’s roadmap or revenue goals — you’re in control.
And for many businesses, that’s the only safe choice.
The Final Moats
If open models are catching up in quality, winning on cost, gaining trust, and accelerating adoption, what’s left for closed providers to hold onto?
Training and running large-scale AI models still requires massive compute power, specialised chips, and significant energy resources. Companies like OpenAI, Anthropic, and Google have access to supercomputers, private data centres, and preferential cloud infrastructure deals. These remain formidable advantages—for now.
The Compute Gap
Training a model like GPT-4 reportedly cost tens of millions of dollars. For most organisations—and certainly most individuals—this kind of scale is out of reach. Closed providers leverage this to justify their prices, protect their IP, and maintain their lead.
But the dynamics are shifting. Open models don’t need to match the biggest proprietary models parameter for parameter. Instead, they win by being good enough and by being optimised for efficiency, not scale.
Models like LLaMA 3, Microsoft’s Phi, Google’s Gemma, and DeepSeek R1 distillations are engineered to run on modest hardware while still delivering state-of-the-art results. Quantisation techniques reduce memory usage. Distillation keeps reasoning power while shrinking in size. And inference frameworks like vLLM and GGUF enable faster generation with lower latency and energy use.
You can now run powerful language models from a laptop—and in many use cases, that’s more than enough. Additionally, the hardware needed to run those models is only going to get cheaper and more accessible.
The Ecosystem
There’s one final frontier that could rival compute as a moat: ecosystem lock-in.
Closed providers don’t just offer models—they offer a tightly integrated stack: chat interfaces, fine-tuning APIs, function calling, embeddings, vector databases, agent frameworks. It’s a whole platform. And if you build your product too closely around it, leaving becomes painful.
But this, too, is changing. Hugging Face, Groq, Fireworks, vLLM, Ollama, Axolotl, LangChain, PostgresML—an ecosystem of open tooling is rising fast, offering modular, interchangeable pieces for every part of the stack.
The result? A slow but steady erosion of the moats closed providers once relied on. When performance is comparable, when tooling is mature, and when deployment is flexible, users will choose the path that offers control, cost savings, and freedom.
Ideological and ethical considerations
Not all arguments for open models are technical or economic. Some are deeply philosophical. As AI becomes more powerful—and more embedded in every aspect of our lives—questions of who gets access, who has control, and who benefits become impossible to ignore.
At the core of this is a simple but profound observation: AI models are trained on humanity’s collective output.
Every blog post, book, comment thread, academic paper, piece of art, and line of code from an open-source project—all of it has contributed to building the models we use today. Our digital footprints, whether volunteered or scraped, have been encoded in those massive neural networks shaping the intelligence now sitting behind APIs and prompts.
The question then is, who should own that intelligence? Who should own and have access to the expression of all humanity encoded in weights inside AI models?
Access vs. Benefits
Companies like OpenAI often say they want to ensure that “the benefits of AI are shared by all.” But there’s a crucial distinction between benefiting from a technology and having access to it.
You can benefit from electricity without generating your own power. You can benefit from medicine without knowing how to make it. But in both cases, you're dependent on others for access, pricing, and availability. In some cases, this trade-off is acceptable, but in others, it is not.
When advanced models are locked behind proprietary APIs—shaped by commercial incentives, usage limits, and opaque filters—access to one of the most transformational technologies ever created, on a par with the invention of personal computers and the internet, is conditional.
Open models flip that dynamic. They offer direct access to the tools themselves, not just the polished products built on top. You can download the model weights, inspect the architecture, fine-tune it to your needs, and run it on your own terms. That’s not just a technical distinction—it’s a different, more democratic relationship with technology. It turns AI from a service you rent into a capability you own.
A force against concentration
Open models also serve as a check on concentrated power.
Just as open-source software prevents a few corporations from full control of the web, open models could prevent a handful of companies from monopolising intelligence. They give governments, educators, nonprofits, and independent developers the tools to compete, innovate, and build without needing a multi-million dollar compute budget or access to a proprietary API.
This matters especially in regions where access to capital is limited, but talent is not. Open-source AI lowers the barrier to entry and allows ideas, not just access to the infrastructure, to be the deciding factor in who gets to build the future.
Yes, there are risks
Of course, open access comes with real risks: misuse, disinformation, and bad actors. These concerns shouldn’t be minimised. But they should also be put in perspective.
Openness doesn’t mean anarchy. It means transparency, peer review, community-led governance, and shared responsibility. These are the same principles that make open-source software more secure, not less.
And if history is any guide, openness is not just a technical choice—it's a moral one. The internet as we know it exists because foundational technologies were shared, not sold. The same could be true for AI.
A fork in the road
The leaked Google memo was a warning to big tech companies: open models are moving faster, costing less, and catching up in capabilities at an astonishing pace. Two years later, that prediction has become a reality.
What was once a fringe movement is now the foundation for the next era of AI. From Llama to DeepSeek R1, from small indie projects running on laptops to large enterprises fine-tuning domain-specific agents, open models are everywhere.
This is more than just a technical shift. It’s a structural one. We are watching AI move from product to platform, from proprietary to participatory, from centralised to distributed. Just like the internet didn’t belong to AOL, and computing didn’t stop with Microsoft, the future of AI will not be owned by a single company.
Closed models will continue to play a role, especially for frontier research and highly advanced tasks. But the centre of gravity is moving. The flexibility, cost-efficiency, transparency, and freedom that open models offer make them more and more too compelling to ignore.
And like every open-source revolution before it, the real winners will be those who understand the shift early—who build, contribute, and grow with the ecosystem, rather than against it. Because the question is no longer whether open models can compete. The question is: what will we build with them?
Thanks for reading. If you enjoyed this post, please click the ❤️ button or share it.
Humanity Redefined sheds light on the bleeding edge of technology and how advancements in AI, robotics, and biotech can usher in abundance, expand humanity's horizons, and redefine what it means to be human.
A big thank you to my paid subscribers, to my Patrons: whmr, Florian, dux, Eric, Preppikoma and Andrew, and to everyone who supports my work on Ko-Fi. Thank you for the support!
My DMs are open to all subscribers. Feel free to drop me a message, share feedback, or just say "hi!"