For a brief period beginning with ChatGPT’s launch in November 2022, the frontier of artificial intelligence looked like it would be owned by a handful of well-capitalised labs: OpenAI, Anthropic, Google DeepMind. The assumption was that only companies with billions of dollars and access to tens of thousands of GPUs could build models worth using. That assumption is now being seriously challenged — not by a single dramatic breakthrough, but by a steady accumulation of open-weight models that are good enough, cheap enough, and free enough to reshape the market.
The Llama Effect
Meta’s decision to release the weights of its Llama models — first as a research-restricted release, then under increasingly permissive licences — was the pivotal moment. Llama 2 in July 2023 was a credible open alternative to models a year older. Llama 3 in April 2024 narrowed the gap further. By Llama 3.1, released in July 2024, Meta shipped a 405-billion-parameter model that approached the performance of GPT-4-class systems on many benchmarks, alongside smaller 8B and 70B versions that could run on commodity hardware.
Llama 3.3 70B, released in December 2024, was arguably the most consequential: a model small enough to run on a single high-end GPU cluster that Meta claimed matched the 405B model on many tasks. Thousands of derivative models — “fine-tunes” — appeared on Hugging Face within weeks, adapted to legal work, medicine, coding, and dozens of languages.
Mistral and the European Challenge
Paris-based Mistral AI became the standard-bearer for European open models. Its Mixtral architecture, which uses a sparse mixture-of-experts design, delivered strong performance at a fraction of the compute cost. Mistral released models under the Apache 2.0 licence, making them genuinely free for commercial use — a meaningful distinction from Meta’s custom community licences, which impose restrictions on very large-scale users. Mistral’s strategy combined open weights with a commercial API, proving that the two could coexist.
China’s DeepSeek Shock
Perhaps the most disruptive development came from China. DeepSeek, a spinout of the quantitative hedge fund High-Flyer, released DeepSeek-V3 in late 2024 — a 671-billion-parameter mixture-of-experts model that its developers said was trained for under $6 million in compute, a figure that stunned observers accustomed to nine-figure training budgets. Then in January 2025, DeepSeek released R1, a reasoning model with open weights that matched or exceeded OpenAI’s o1 on several benchmarks.
The market reaction was immediate and dramatic: Nvidia lost hundreds of billions of dollars in market value in a single day as investors questioned the assumption that frontier capability required frontier-scale spending. Whether DeepSeek’s cost claims were fully apples-to-apples is debated, but the strategic point stood. Capable models could be built more cheaply than the incumbents’ business models assumed.
The Benchmark Reality
On standardised evaluations such as MMLU, GPQA, and HumanEval, the gap between the best open-weight models and the best proprietary models has shrunk from two years to a matter of months. On some tasks — notably coding and mathematical reasoning — open models are competitive with or ahead of proprietary systems. On others, particularly long-context reasoning, tool use, and multimodal tasks, the frontier labs retain an edge.
Benchmarks are imperfect measures, and the modern practice of training on benchmark-adjacent data muddies comparisons. But directionally, the trend is unmistakable: the open ecosystem is roughly six to twelve months behind the frontier, and closing.
Why This Matters Economically
The implications extend well beyond benchmarks. When a capable model is free to download, the marginal cost of inference becomes a matter of hardware and electricity, not API pricing. Companies with privacy, latency, or cost requirements can run models on their own infrastructure. Startups can build without paying per-token fees that scale with success. Regulated industries can keep data in-house.
This creates a two-tier market. Frontier labs compete on capability at the very top — the most complex reasoning, the largest context windows, the best multimodal understanding — and charge premium prices. The open ecosystem commoditises everything below that frontier, driving down prices and forcing proprietary providers to justify their premiums with real differentiation.
The Cloud Doesn’t Lose — It Adapts
Notably, the hyperscalers themselves are embracing open models. AWS, Google Cloud, and Microsoft Azure all offer hosted versions of Llama, Mistral, and other open models alongside proprietary ones. Their business is compute and distribution, not model ownership per se. If open models win, the cloud still wins — it just sells GPU hours instead of tokens.
Looking Forward
The open-source AI movement faces real headwinds. Training frontier models still requires enormous capital, and the best models may always arrive first behind closed doors. Safety debates — whether releasing powerful weights is wise — continue. And Meta’s willingness to keep releasing frontier-scale weights is a strategic choice, not a guarantee.
But the direction of travel is clear. The idea that artificial intelligence would be a moat controlled by a few firms has given way to something more like the Linux story: a shared foundation that many companies build upon, with proprietary value created at the edges. For Canadian enterprises in particular — from Toronto fintechs to Montreal AI labs — that shift lowers the barrier to deploying serious AI without signing away their data or their margins.
The Licensing Slippage
“Open source” in AI is a slippery term, and the details matter. Meta’s Llama models are open-weight, but the licence carries restrictions — the Llama Community Licence requires a separate agreement for companies with more than 700 million monthly active users, and it imposes acceptable-use conditions. It is not the same as the OSI-approved copyleft of Ghostscript or the permissive Apache licence of Mistral’s models. Purists argue, correctly, that this makes Llama “open weight” rather than “open source.” For businesses, though, the practical question is narrower: can we download it, run it, fine-tune it, and deploy it without paying per token? For almost all of them, the answer is yes.
The distinction matters for the long run. Licences can be changed — and Meta, crucially, has reserved the right to do so for future models. A company that builds its entire product on a third-party foundation is exposed to the licensor’s decisions. This is why some enterprises favour Apache-licensed models, or invest in developing their own sovereign capabilities.
Sovereignty and the Data Question
For governments and regulated industries, the open-weight movement is fundamentally about data sovereignty. A hospital cannot send patient records to a foreign cloud API without running afoul of privacy law and ethical obligations. A defence agency cannot route sensitive queries through a commercial service. An open model that runs inside the organisation’s own data centre solves these problems in a way no API can. This is driving public investment: countries and blocs are funding sovereign AI programmes and national compute, viewing dependence on foreign frontier models as a strategic vulnerability.
The Economics of Inference
The cost structure of AI is asymmetric, and open models exploit the asymmetry. Training a frontier model is a fixed, enormous cost paid once. Inference — answering queries — is a variable cost paid forever, and it dominates lifetime expenditure for widely-used systems. If an organisation can avoid per-token pricing by running an open model on its own or rented hardware, it captures the entire margin that an API provider would otherwise take. For high-volume applications, the savings are transformative. This economic logic, more than ideology, is what drives open-weight adoption in production.
The Fine-Tuning and Distillation Flywheel
The most underrated advantage of open weights is the ability to specialise. A general model is a jack of all trades; a fine-tuned model can be dramatically better at a narrow task. Companies take a base open model and train it on their own domain data — legal precedent, medical guidelines, proprietary code — producing a specialist that outperforms a far larger general model on their task while costing a fraction to run. Distillation, in which a large model’s outputs train a smaller one, compounds this: capable small models can be created cheaply. This flywheel is why the open ecosystem produces so many useful derivatives so quickly.
Safety, Misuse, and the Genie Problem
Releasing weights means giving up control. A closed API can refuse harmful requests, log abuse, and shut off access. An open model, once downloaded, cannot be recalled. Critics argue this makes powerful open models a proliferation risk, particularly for misuse in bioweapon design, cyberattacks, or disinformation. Defenders counter that the capability gap between the best open and closed models is not large enough to matter as a safeguard, that openness enables defensive research and independent audit, and that centralised control creates its own dangers of concentrated power. The debate is unresolved, and it is genuinely difficult: both positions have real weight.
Benchmarks, Leaderboards, and Their Limits
The open-versus-closed debate is frequently conducted through benchmarks, and benchmarks deserve scrutiny. Leaderboards like MMLU, GPQA, and LMSYS Chatbot Arena provide useful snapshots, but they are imperfect. Training on benchmark-adjacent data can inflate scores. Different evaluation conditions — temperature settings, prompting, tool use — make comparisons unreliable. And for real-world tasks, raw benchmark performance often matters less than reliability, latency, cost, and integration quality. A model that scores 2% lower on a multiple-choice exam but runs four times faster and costs a tenth as much may be the better product. The practical competition is not for a leaderboard position but for the workflow in which the model is embedded.
