Rethinking Openness in the Age of Foundation Models
As the global AI community wrestles with the tension between scale and accessibility, recent work by researchers at Lelapa AI offers a compelling example of what inclusive, efficient AI can look like. Around the world, the push to shrink massive language models into deployable, localized tools has become a strategic priority, whether to reduce energy costs, mitigate cloud dependency, or bring AI to underserved populations. Yet, few efforts embody this shift as clearly as the recent transformation of InkubaLM.
Last month, seven African researchers shrank a multilingual language model by 75 percent—an outcome made possible not by brute compute but by strategic compression and the enabling role of open-source architecture. The feat happened during Lelapa AI’s Buzuzu-Mavi Challenge, where 490 participants from 61 countries were invited to compress InkubaLM, Africa’s first multilingual Small Language Model (SLM) for five African languages. Built from scratch, according to its model card on Hugging Face, InkubaLM follows a lightweight adaptation of LLaMA-7B’s open-source architectural design, optimized for low-bandwidth environments.
By making the model architecture and weights openly available, Lelapa AI empowered participants to inspect, adapt, and optimize the model without proprietary restrictions. The top three teams, all from Africa, employed a combination of open-source techniques—distillation, quantization, vocabulary pruning, adapter heads, and shared embeddings—to reduce the model's parameters to just 40 million while maintaining translation quality for Swahili and Hausa.
In a continent where only one-third of people have reliable internet and 70 percent rely on entry-level smartphones, compressing a model like InkubaLM to run offline marks a turning point. It's the difference between AI locked behind a cloud subscription and AI that can function on a low-cost smartphone in Kisumu or Kano. Because InkubaLM was released under an open license and built on an inspectable, adaptable architecture, African researchers were able to tailor it for their realities. That openness at the architectural and licensing levels was essential to enabling offline deployment and real-world relevance.
Framing the Question
The episode highlights a broader debate reverberating across the AI world: Should the power to build and adapt AI remain concentrated in a handful of companies, or be distributed to anyone with sufficient computing power and access to open tools? This distinction has far-reaching implications—not just for innovation, but for who gets to shape AI in their own context. To answer, we first need to unpack what “open-source AI” really means, why the term is suddenly contested, and how the battle lines—risk containment versus democratic access—shape everything from export-control law to whether a farmer in Kisumu can receive drought advice in Kiswahili.
What Is Open-Source AI?
Defining 'open-source AI' has become surprisingly contentious, precisely because AI systems don't fit neatly into traditional open-source software categories. The classic open-source definition, originating from software, centers on the freedom to use, study, modify, and share code. But AI introduces new components that complicate this framework: model weights (the trained parameters), training data (often massive and sensitive), and computational requirements that can make 'modification' practically impossible for most users.
Some advocates argue that truly open-source AI must include all three components: source code, model weights, and training data (or transparent documentation of training methods). Others contend that releasing weights and code is sufficient, even if the training data remains proprietary. Still others question whether the open-source framework applies at all—after all, you can't easily 'read' billions of floating-point parameters the way you can read a Python function.
In practice, most current AI releases only provide some of these pieces. It's common to see model weights made public, but with the training data and methodology kept private. This halfway approach, sometimes criticized as openwashing, has created widespread confusion. Are we examining a truly open-source model, or merely a publicly downloadable artifact?
The Open Source Initiative has now drawn a definitive line with their official Open Source AI Definition: AI systems must include detailed data information, complete source code, and model parameters to qualify as open source. If you can't inspect, reproduce, and modify all three components, it isn't open source. By this standard, most AI systems currently labeled as "open source", including popular models like LLaMA 2 and many others, fall short. Anything less might still be helpful, but it lacks the transparency and collaborative potential that the term "open-source" implies.
Risk Containment vs. Democratic Access
The debate over open-source AI is often framed as a security tradeoff: does transparency increase risk, or reduce it through collective oversight? That fault line has now hardened into two major camps.
Camp 1: Closed = Safer
This group comprises most of the major labs (OpenAI, Google DeepMind, Anthropic) and a number of national security experts. Their view is that advanced AI is a dual-use technology, similar to nuclear research or bioengineering. They argue that open-sourcing powerful models too early could let anyone, even those without specialized knowledge, cause significant harm.
OpenAI, for example, initially committed to openness but now publishes far less content than it once did. This more cautious stance has gained traction across the field, from Anthropic’s warnings about AI-generated bioweapons to former Google CEO Eric Schmidt’s calls for export controls on AI models. Underlying these positions is a growing concern that transparency, rather than preventing harm, may actually accelerate misuse.
In response, these organizations typically favor managed access, offering their models via APIs, gated licenses, or safety filters that restrict use to approved contexts. This helps them monitor deployments, enforce safeguards, and intervene when necessary. It also, not incidentally, aligns with their commercial interests.
Camp 2: Open = Safer
On the other side are open-source communities like LAION, AI startups like Mistral, and major researchers like Meta's Yann LeCun. They argue that openness breeds trust, improves safety, and decentralizes power.
This camp often draws inspiration from the broader open-source software movement. The Transformer architecture, PyTorch, and even early versions of GPT were openly published, spurring a flood of innovation. In this view, AI advances more rapidly and becomes safer when more people can inspect, improve, and adapt it.
For example, Meta's LLaMA 2 and Mistral's 7B models were released with permissive licenses, and the response was immediate: within weeks, the community had developed safety filters, plugins, fine-tunes, and other adaptations. These were improvements that no single company, however well-funded, could have accomplished alone.
But commercial motivations complicate this narrative. Mistral aggressively lobbied EU lawmakers to abandon plans for regulating foundation models directly, arguing instead that only AI applications should face "hard rules." CEO Arthur Mensch framed this as protecting European competitiveness, telling lawmakers that foundation models are "nothing else than a higher abstraction to programming languages" and shouldn't be regulated differently. Critics noted that this position, shared by U.S. giants like Microsoft and OpenAI, would conveniently exempt Mistral's business model while burdening downstream competitors. Just months after securing favorable treatment in the EU AI Act, Mistral announced its partnership with Microsoft, leading observers to question whether its open-source advocacy was genuine principle or strategic positioning.
Similarly, Meta's embrace of openness coincided with their struggle to compete directly with ChatGPT; releasing LLaMA 2 helped them commoditize their competitors' advantages while building an ecosystem around their own architecture.
Still, regardless of motivation, the technical benefits remain real. Open source AI models allow customization for local languages (as seen in the case of InkubaLM), cultures, and needs—something closed models can't do effectively. Advocates view this as a means to challenge monopolies, expand access, and foster a more equitable AI future.
What’s at Stake
These camps and their relative perspectives reflect deeper values about how AI should be governed and by whom.
Trust and Transparency: If people don’t understand how AI systems are trained or what data went into them, it’s hard to build trust. Open models offer a pathway to transparency and accountability.
Governance: Closed models consolidate control in a handful of companies. Open models invite collaborative oversight, which may be essential for democratic governance of powerful technologies. They also support the development of AI as digital public infrastructure (DPI): Tools that are inspectable, auditable, and governed in the public interest, much like open standards for identity or payments.
Innovation: Openness accelerates progress. Instead of duplicating work behind closed doors, open-source fosters a shared foundation for experimentation and invention.
Equity: Who gets access to AI? Just the well-funded firms in Silicon Valley or researchers, nonprofits, and startups around the world? Openness lowers the barrier to entry and levels the playing field. It also enables countries and communities to build and adapt AI systems without relying on closed, foreign-controlled platforms. This is an essential step toward achieving digital sovereignty and an inclusive digital public infrastructure (DPI). But openness at the model level doesn't guarantee equitable access if other layers remain concentrated. Even with open weights, organizations still need substantial computing resources, specialized expertise, and platform infrastructure, creating new chokepoints where power can reconcentrate. True equity requires addressing these deeper dependencies, not just open-sourcing the models themselves.
Security: There’s no easy answer here. Open models could be misused, but they also enable more people to build defenses, audit vulnerabilities, and reduce reliance on opaque systems.
The Gray Area: Openness Is a Spectrum
Most "open" AI models today occupy a problematic gray area, complicating efforts to define and defend genuine openness. For instance, Meta’s widely celebrated LLaMA 2 made code and model weights available but withheld critical training data and included restrictive licensing terms, limiting practical reuse. Similarly, Mistral 7B took a more permissive stance but still fell short of full transparency, offering limited insight into its dataset and training methodologies.
Even Stable Diffusion, frequently praised as a landmark example of openness, was trained on datasets compiled from scraped online images without explicit consent, raising substantial privacy and ethical concerns. This opaque data sourcing approach undermines the foundational principles of transparency and accountability that open-source advocates champion.
These ambiguities create genuine uncertainty for policymakers, developers, and end-users. When the exact boundaries of openness remain blurred, it becomes challenging to establish clear governance standards, assess security implications accurately, and ensure responsible use. In short, partial openness isn't merely incomplete—it risks undermining trust and diluting the very benefits that open-source AI promises to deliver.
Our Take, And Why This Matters
Foundation LLMs are powerful. They provide organizations with quick access to advanced language capabilities, enabling teams to accelerate the development of tools for tasks such as summarization, classification, and information retrieval. For many, these large pre-trained models offer a fast on-ramp to experimenting with generative AI.
But smaller open-source models (often referred to as Small Language Models, or SLMs) offer a different kind of opportunity. They allow organizations to move beyond consumption and begin shaping AI to reflect their own needs, fine-tuning for specific tasks, local languages, cultural contexts, or sector-specific domains. Open-source models unlock a layer of ownership that closed systems rarely allow.
This flexibility is especially important for organizations with both technical capacity and specialized requirements that commercial APIs can't address - research institutions, government agencies with unique security needs, technology companies building AI-powered products, and organizations working in underserved languages or domains. For teams with the necessary expertise, being able to start from a well-built, downloadable model opens up possibilities that proprietary APIs simply can't match: experimentation in secure environments, deep customization, and complete control over deployment.
And increasingly, these decisions carry geopolitical weight. As Aubra Anthony notes in Lawfare, many countries, particularly in the Global South, aren’t seeking AI sovereignty out of pride, but to avoid being locked into systems that don’t serve their languages, cultures, or priorities. Open models offer a more realistic path to autonomy, providing the ability to shape and govern AI tools without needing to own the entire compute stack.
In this sense, open-source LLMs aren’t just customization tools; they are a form of digital public infrastructure. Like open standards in payments or broadband, they provide a shared, inspectable foundation that others can build on without surrendering control. This framing matters not only for countries seeking technological autonomy, but also for nonprofits, civic institutions, and small businesses here in the U.S., many of whom face the same pressures: rising API costs, vendor lock-in, and limited influence over the tools they rely on. Recasting openness as DPI clarifies what’s at stake, not just convenience or cost, but sovereignty, resilience, and meaningful participation in the AI era.
Final Reflections
InkubaLM's success shows what's possible when communities shape AI on their own terms. But the broader lesson is about competition and choice. When a handful of companies control the foundational infrastructure of AI—from training to deployment—they effectively set the terms for everyone else's innovation.
The risk is market concentration that stifles the kind of experimentation that led to InkubaLM's breakthrough. If foundation models remain locked behind proprietary APIs, if regulatory frameworks favor incumbent platforms, and if the compute requirements for meaningful AI development stay prohibitively high, we're cementing a handful of companies as permanent gatekeepers.
Open models don't guarantee better outcomes, but they preserve competitive pressure. They force proprietary providers to compete on performance rather than locking them in. They enable new entrants to build on shared foundations rather than having to start from scratch. They ensure that innovation can occur outside the strategic priorities of a few dominant firms.
But openness alone isn't sufficient. As Nadia Andrada recently noted in ICTworks, organizations in the development sector that built everything on open-source principles while remaining grant-dependent found themselves vulnerable when funding disappeared, even as commercial companies monetized their open-source work. Pure open-source approaches without viable business models can actually reduce access over time, leaving promising technologies unused while commercial alternatives thrive.
With this in mind, the question ahead is whether we can build sustainable ecosystems around open models that preserve competitive pressure and democratic participation while ensuring the infrastructure remains viable and sustainable. Without this competitive dynamic, we risk outsourcing not just tools, but the entire direction of AI development to the internal roadmaps of a few powerful companies.
The stakes are significant, but the reality is messier than a simple choice between open and closed futures. Market competition, democratic input, and corporate strategy will all shape how AI develops, often in contradictory ways simultaneously. The question isn't whether alternatives will remain possible; they will. It's whether they'll remain practical and competitive enough to provide meaningful choice and constraint on dominant platforms. InkubaLM's success shows that innovation can still emerge, but sustaining alternatives requires more than just technical capability. It requires viable ecosystems that can compete with entrenched platforms over time, while navigating the same economic and political pressures that shape all technology markets.
