The Open-Weight AI War: How Llama Opened the Door and DeepSeek, Qwen, and Mistral Changed the Market

One Billion, Then the Flood

On March 18, 2025, Mark Zuckerberg posted a clean platform metric: Llama had crossed 1 billion downloads.

It was a perfect Meta number. Large. Clean. Hard to argue with. It suggested that the old debate was over. Powerful AI no longer had to be rented by the API call. Developers could download it, fine-tune it, and run it on their own terms.

But the number also arrived a little late.

By the time Meta reached that symbolic threshold, the open model story no longer belonged to one company. DeepSeek had shown that reasoning itself could escape the closed labs and become something engineers could download, distill, and tune. Mistral had turned permissive licensing into a business argument for Europe. Alibaba’s Qwen line had begun to look less like a single model family and more like a release machine.

Hugging Face’s own late-2025 analysis captured the shift. In the 1B+ parameter segment, Meta still led total downloads with 23.2%. Alibaba had already reached 20%. Mistral held 6.8%. DeepSeek had 3.8%.

Another 15.6% of downloads in the same segment were captured not by the original labs at all, but by individuals who packaged quantized variants for the community. That number mattered as much as Meta’s lead. It meant the market was already slipping out of the hands of the original publishers.

The real story of open AI in 2025 and early 2026 is not “open versus closed.” It is a struggle between several different ideas of what openness is for.

Meta uses open weights to extend platform influence. Mistral uses them to sell trust and sovereignty. DeepSeek uses them to compress the reasoning gap and reset cost expectations. Qwen uses them to spread across languages, clouds, and hardware tiers.

Buyers are asking different questions now. Not just which model tops a benchmark, but which one legal will approve, which one infra can run locally, which one procurement can defend, and which one can be adapted without asking a vendor for permission.

That last question is not academic. It shapes which models get built into products, which vendors survive compliance review, which governments subsidize local champions, and which ecosystems become defaults. The past two years did not simply broaden access to AI. They turned openness into a strategic market design choice.

Meta Built the Market, Then Wrapped It in Conditions

Meta deserves credit for one fact that many of its critics understate: Llama changed developer behavior before any other major company managed to do it at comparable scale.

The breakthrough was not that Meta released weights. Smaller labs had done that before. The breakthrough was repetition. Meta kept releasing models that were useful enough, large enough, and visible enough to shape the rest of the market. Llama was never just a model family. It became a coordination device. People built serving stacks, quantization pipelines, fine-tuning recipes, and local tools around it. Without saying so explicitly, they standardized around Meta’s default.

That substrate grew more ambitious in April 2025 with Llama 4. Meta’s published model card described Scout as a 17 billion active-parameter model with 109 billion total parameters and a 10 million-token context window. Maverick kept the same 17 billion activated parameters but expanded to 400 billion total parameters with a 1 million-token context window.

Meta said Scout was pretrained on roughly 40 trillion tokens and Maverick on roughly 22 trillion. The pretraining bill, by Meta’s own accounting, totaled 7.38 million H100 GPU-hours. The company also reported 1,999 tons of location-based greenhouse gas emissions for the two model trainings combined.

These were not “community project” numbers. They were industrial numbers.

The training data disclosure made the structure even clearer. Meta said Llama 4 was trained on a mix of publicly available and licensed data plus information from Meta’s own products and services, including publicly shared Facebook and Instagram posts and people’s interactions with Meta AI. This is the heart of Meta’s asymmetry. The weights are downloadable. The data engine that made them is not. The open layer sits downstream from one of the world’s largest private reservoirs of human interaction data.

That distinction matters because it explains what Meta actually wants from openness. The company does not need Llama to be fully open in the classic software sense. It needs Llama to be open enough that the rest of the market builds around Meta’s formats, tooling, and assumptions.

The license confirms that intent.

Llama 4 is not released under Apache 2.0 or MIT. It comes under the Llama 4 Community License, a custom commercial license. Companies with more than 700 million monthly active users need a separate grant. Redistributors must carry Meta’s license text. Products or documentation must display “Built with Llama.” If someone uses Llama materials or outputs to create a distributed derivative model, Meta requires “Llama” at the beginning of the model name.

These clauses do not make Llama unusable. They make it governed.

The Open Source Initiative has been blunt about this. In the Open Source AI Definition published in October 2024, OSI argued that genuinely open-source AI requires not only weights and architecture but also the data information and code used to derive those weights. In February 2025, OSI wrote that Meta’s newer Llama licenses still failed the open-source test and continued to restrict who could use the models and under what conditions. Whether one agrees with OSI’s full position or not, the dispute is economically important. If openness is attached to legal conditions, brand control, and undisclosed upstream data advantages, then “open” is functioning less like public infrastructure and more like strategic ecosystem expansion.

Meta is comfortable with that ambiguity because it works in the company’s favor.

The consumer layer reinforces it. The standalone Meta AI app launched in April 2025 built directly on Llama 4 and tied together Meta’s app surfaces, glasses, and web experience. The company could tell developers it believed in open AI while simultaneously using the same model family to deepen its closed product loop across WhatsApp, Instagram, Facebook, Messenger, and hardware. That is not a contradiction. It is the business model.

The paradox is that Meta’s success taught the rest of the market to ask questions that weaken Meta’s unique advantage. Once enterprises accept that powerful AI can be downloaded, inspected, and self-hosted, they start comparing licenses, governance terms, hardware footprints, and regional restrictions. On those dimensions, Meta no longer looks like the obvious default.

Llama opened the market. It also trained the market to shop around.

Mistral Turned Permissive Licensing Into a Business Model

Arthur Mensch’s company took almost the opposite path.

Where Meta used open weights to extend platform power, Mistral used permissive licensing to manufacture trust.

That difference stops sounding philosophical once the releases are laid side by side.

In January 2025, Mistral introduced Mistral Small 3, a latency-optimized 24B model under Apache 2.0, and pitched it as an open replacement for opaque proprietary small models. The company said it scored above 81% on MMLU while running at around 150 tokens per second and more than three times faster than larger alternatives on the same hardware.

In March it followed with Mistral Small 3.1, again under Apache 2.0, with stronger text performance, multimodal support, 128k context, and enough efficiency to run on a single RTX 4090 or a Mac with 32GB RAM. In June came Magistral, the company’s first reasoning line, released in both open and enterprise variants. In July came Voxtral, an open speech-understanding family, also under Apache 2.0.

The pattern was not accidental. Mistral was not dabbling in openness. It was productizing it.

That mattered because the market had already grown suspicious of model vendors that used the language of community while tightening control the moment enterprise money appeared. Mistral did something rarer. It kept treating permissive release as part of the offer. The open models were not a funnel to something else. They were the foundation of the something else.

This is how the company separated itself from both the American hyperscalers and the Chinese labs. For a European buyer, Apache 2.0 is not a cultural flourish. It is a way to make a procurement process move.

In many organizations, the real comparison does not happen on X or on a benchmark site. It happens in a meeting with product, security, infra, and legal. If legal has to escalate the license, the pilot slows down. If the model can be self-hosted and modified under familiar terms, the meeting gets shorter. That is a real advantage.

Mistral built a business on top of that convenience.

Le Chat, the company’s consumer assistant, crossed 1 million downloads in roughly two weeks after the February 2025 mobile launch. Emmanuel Macron publicly urged French viewers to download it ahead of the Paris AI Action Summit, a small but revealing moment. Mistral was no longer simply a Paris research startup. It had become a symbolic instrument in Europe’s effort to prove it could participate in frontier AI without surrendering the entire stack to U.S. platforms.

The enterprise layer moved faster than the consumer layer. In May 2025 Mistral launched Le Chat Enterprise with enterprise search, agent builders, custom data and tool connectors, document libraries, custom models, and hybrid deployment.

The same month it made the commercial logic explicit: open models below, managed enterprise workflows above. The partnership with AFP added another piece. AFP said Le Chat would gain access to its daily stream of 2,300 text stories across six languages. That was not just a content deal. It was Mistral answering the usual enterprise complaint about open models: yes, they are powerful, but who makes them dependable?

The financial and industrial context made the bet more credible. In 2025 Mistral’s valuation rose sharply, and by early 2026 Arthur Mensch was publicly saying the company should exceed 1 billion euros in revenue during 2026, while also telling the Financial Times that annualized run-rate revenue was already north of $400 million. At the same time, the company committed 1.2 billion euros to data centers in Sweden, signaling that “European sovereign AI” would eventually require European compute, not just European branding.

The skepticism is straightforward. Mistral still operates far below the capital scale of Meta, Google, or OpenAI. If the cost of frontier training continues to rise, the temptation to close more of the stack will grow. Open models are expensive to maintain. Support is expensive. Cloud bills do not care about ideology.

But so far Mistral has held a coherent line. It has been more permissive than Meta, more enterprise legible than DeepSeek, and more politically usable than many American labs. Its strategic claim is not that open models should replace every closed service. It is that open models plus a strong enterprise layer can become the preferred option wherever trust, inspectability, and sovereignty matter more than mass consumer distribution.

That is a narrower claim than Silicon Valley usually makes. It may also be a more defensible one.

DeepSeek Made Reasoning Portable

If Meta opened the market and Mistral gave one part of it a business model, DeepSeek changed the thing people thought the market could buy.

Before DeepSeek-R1, the most important feature still thought to be meaningfully protected by closed labs was reasoning.

That was the core of the o1 moment. It was not just that OpenAI had a strong model. It was that it appeared to have found a way to make inference-time compute itself into a commercial moat. The model could “think” longer, reason more carefully, and solve harder problems by spending additional computational budget during response generation. For much of 2024, that looked like something you would rent through a closed API, not something you would download.

DeepSeek attacked exactly that assumption.

The January 2025 R1 release was technically strong enough to matter. What made it historic was the package around it.

DeepSeek did not just ship weights. It shipped a technical narrative that other engineers could work with. Hugging Face’s Open-R1 team later described DeepSeek-V3, the base model behind R1, as a 671B mixture-of-experts system performing alongside major closed models while claiming a training cost of about $5.5 million, thanks to a combination of architectural choices and hardware optimization. DeepSeek-R1 itself showed how staged reinforcement learning and post-training could generate high-level reasoning behavior outside the closed-model club.

Then the release started mutating.

DeepSeek published distilled models built on both Qwen and Llama backbones. That single decision mattered almost as much as the flagship model did. It told the market that advanced reasoning behavior could be ported across open families rather than remaining locked to one lab’s architecture. Reasoning stopped looking like a one-company property and started looking like an engineering layer.

Hugging Face understood the implication immediately. Its Open-R1 project, launched days later, aimed to reconstruct the missing pieces of DeepSeek’s pipeline so the broader community could reproduce and extend it. In January 2026 Hugging Face would look back and argue that R1 lowered three barriers at once: the technical barrier to reasoning, the adoption barrier to deploying it, and the psychological barrier around who could lead in open AI.

The psychological shift was real. DeepSeek-R1 became the most liked model on Hugging Face. It also forced Western labs to respond strategically, not rhetorically. Meta reportedly organized internal “war rooms” to study DeepSeek’s techniques. OpenAI accelerated its own open-weight posture later in the year. A release from Hangzhou had changed planning cycles in Menlo Park and San Francisco.

But DeepSeek’s story is often flattened into myth, and the myth obscures the more interesting truth.

The first myth is that R1 made frontier AI cheap. It did not. The famous sub-$6 million number, as several analysts later emphasized, captured a narrow training slice rather than the full historical cost of hardware acquisition, experimentation, and infrastructure. CNBC cited estimates that DeepSeek’s broader hardware spend over time could be far higher than the viral headline implied.

R1 did not abolish capital intensity. It embarrassed some assumptions about it.

The second myth is that R1 was fully open source. It was not. Hugging Face launched Open-R1 precisely because important parts of the full training process, especially datasets and code, remained unavailable. DeepSeek was more open than most frontier labs in the parts that mattered most to community progress. It was not transparent end to end.

That is what made R1 matter. It occupied the exact middle ground that changes industries: open enough to copy, incomplete enough to provoke reconstruction.

There was also a geopolitical dimension. DeepSeek’s release gave Chinese open AI a template that could be extended faster under constraint. Hugging Face’s January 2026 retrospective argued that after R1, Chinese companies moved from isolated model launches toward a broader ecosystem strategy. Releases became more frequent. New organizations entered the field. Open models from China began to dominate weekly popularity among newly created repositories. DeepSeek was not the whole ecosystem, but it was the inflection point that made the ecosystem look self-reinforcing.

Liang Wenfeng’s company, in that sense, did something that many better-funded labs failed to do. It made openness contagious.

Once reasoning became portable, it could be quantized, specialized, and deployed. That changed the question from “who has the secret sauce?” to “who can absorb the sauce fastest?” Closed labs still had stronger products, deeper integrations, and more polish. But one of their most important technical advantages had become a downloadable asset.

That is a very different kind of market.

Qwen Industrialized Open Release

The company best positioned to capitalize on that new market was not necessarily the one that created the shock. It was the one that knew how to manufacture follow-through.

That company was Alibaba.

Qwen’s great advantage is not a single signature release. It is cadence, breadth, and system design.

When Alibaba introduced Qwen3 in April 2025, it did not present one heroic flagship and ask the world to admire it. It shipped two MoE models and six dense models at once.

The flagship Qwen3-235B-A22B carried 235 billion total parameters with 22 billion activated. The smaller MoE used 30 billion total with 3 billion activated. Dense variants ranged from 0.6B to 32B. The whole family was released under Apache 2.0.

The official post emphasized support for 119 languages and dialects and roughly 36 trillion training tokens, nearly double the 18 trillion tokens used in Qwen2.5. It also recommended practical deployment tools such as Ollama, LMStudio, MLX, llama.cpp, KTransformers, SGLang, and vLLM. The point was obvious. This was not just a model launch. It was a deployment package.

That combination is easy to underrate from Silicon Valley because it does not fit the frontier-lab theater model.

Qwen is not selling drama. It is selling coverage.

The multilingual strategy is the clearest example. Western labs often treat language expansion as a second-step feature. Qwen treats it as a default market assumption. Supporting 119 languages and dialects is not only a benchmark flex. It changes the addressable market for developers in Asia, the Middle East, Africa, Eastern Europe, and multilingual enterprise environments. It also accelerates local fine-tuning loops, because the base model is already closer to the use case before anybody begins adaptation.

By January 2026, Alibaba said Qwen had crossed 700 million downloads on Hugging Face. Even if download counts are an imperfect proxy for production deployment, they reveal a powerful shift in developer attention. Late-2025 Hugging Face analysis already suggested that Alibaba was on track to overtake Meta in the 1B+ open-model segment if Meta did not refresh its open releases quickly enough. In the 7.5B+ range, Alibaba had already moved ahead.

The more important insight, though, came from Hugging Face’s broader ecosystem review. The platform argued that after the DeepSeek moment, Chinese companies did not simply produce stronger models. They produced more of them, more often, and across more organizations. Baidu moved from no Hugging Face releases in 2024 to more than 100 in 2025. Moonshot’s Kimi K2 became, in the words of some open-model observers, “another DeepSeek moment.” Qwen and Zhipu expanded beyond publishing weights into building interfaces, tooling, and engineering systems around them. Competition, Hugging Face argued, had shifted from raw model comparison toward ecosystems, application scenarios, and infrastructure.

Alibaba is structurally built for that kind of competition.

It has a cloud business. It has enterprise relationships. It has incentive to spread model adoption even when direct monetization is incomplete, because model adoption drives infrastructure demand and ecosystem lock-in elsewhere. Meta’s open strategy is tied to consumer platform gravity. Alibaba’s open strategy is tied to industrial coverage. The difference matters. One tries to make the ecosystem orbit a social platform. The other tries to make the ecosystem run through a cloud and tooling fabric.

This also explains why Qwen is strategically more dangerous than its Western coverage often suggests. It is not just a model family. It is a release machine. If the open AI market rewards breadth, frequency, multilingual range, and deployment convenience, Alibaba may be closer to the market’s center of gravity than many American observers are comfortable admitting.

The Middlemen Became Kingmakers

Closed-model companies like to imagine that the important relationship in AI is vendor to user.

Open-weight AI works differently.

In this market, the middlemen matter almost as much as the original labs. Hugging Face. Quantizers. Ollama. LM Studio. vLLM. cloud marketplaces. inference providers. open-source framework maintainers. These are not distribution afterthoughts. They are part of the product.

This is one reason raw download numbers can be misleading. Hugging Face’s late-2025 analysis of the 1B+ segment found that 15.6% of downloads were being captured by individuals who uploaded quantized versions of base models rather than by the original creators. That is a remarkable leakage of value. It means that in open AI, part of the market advantage goes to whoever makes a model easiest to compress, repackage, serve, and run cheaply on real hardware.

The labs know this, which is why so many recent releases now read like infrastructure manuals as much as model announcements.

Mistral’s Small 3.1 release stressed that the model could run on a single RTX 4090 or a Mac with 32GB RAM and was available via Hugging Face, Mistral’s own platform, and Google Cloud Vertex AI. Qwen’s Qwen3 announcement explicitly recommended local runtimes and production-serving frameworks from day one. Meta’s Llama ecosystem has benefited for two years from exactly this kind of intermediary reinforcement, even when the company itself did not control the packagers doing the work.

This distribution logic changes competitive dynamics in subtle ways.

First, the easiest model to adopt often beats the most elegant model to admire. A team choosing a base model for a private coding agent or multilingual support bot does not always need the absolute top score on a leaderboard. It needs a model that is already packaged for the right runtime, available with the right quantization, documented for the right framework, and safe enough to clear internal review.

Second, permissive licenses compound through intermediaries faster than restrictive ones. Apache 2.0 and MIT make it easier for third parties to wrap, ship, and commercialize deployment tooling without consulting lawyers first. Custom licenses can still win, but they introduce friction precisely in the places where open ecosystems scale: repackaging, naming, redistribution, and integration.

Third, open-weight markets do not fully belong to model creators. They belong to networks of adapters. This is where Meta’s 1 billion download milestone becomes less absolute than it first appears. Llama seeded the ecosystem, yes. But much of the ecosystem’s daily velocity now lives in the hands of packagers, serving-tool builders, cloud partners, and downstream fine-tuners. The original lab retains influence. It does not retain monopoly control.

That is why the open model leaders increasingly look like infrastructure strategists rather than pure research organizations. They are not just trying to train better systems. They are trying to make those systems travel.

The Benchmark Table No Longer Tells You Enough

For much of 2024, model comparison still resembled flagship-phone comparison. Bigger launch, cleaner chart, better top-line number.

That is not how most serious deployments are decided anymore.

The questions people ask in 2026 are more practical and more revealing:

Is the license clean enough for legal to approve?
Can we run this on our own infrastructure?
Does the model already support the languages we need?
Can we fine-tune it cheaply?
Will cloud partners host it?
How much engineering work is required before it feels production-ready?
If we build on top of it, who controls the branding, redistribution, or upgrade path?

Seen through that lens, the open model market looks less like a leaderboard and more like a matrix of trade-offs:

Model family	What is open	License reality	Core strength	Strategic weakness
Meta Llama	Weights and broad ecosystem access	Custom community license with scale and branding conditions	Massive installed base, multimodal capability, tooling inertia	Not fully open-source by OSI standards; ecosystem control remains centralized
Mistral	Broad model release under Apache 2.0	Most permissive among major labs	Enterprise legibility, edge deployment, European trust	Far less capital than U.S. hyperscalers
DeepSeek	Weights and unusually rich technical disclosure	More permissive than most peers, but not full transparency	Reasoning efficiency, cost pressure, derivative velocity	Cost claims are often overstated; geopolitics complicates adoption
Qwen	Multi-size families under Apache 2.0	Highly permissive and deployment-friendly	Multilingual breadth, release cadence, cloud fit	Trust and compliance still vary by market

Put a product manager, an infrastructure lead, a security reviewer, and a lawyer in one room and the discussion stops sounding like an arena leaderboard. The lawyer asks about the license. Infra asks whether the model fits the available GPUs. Security asks whether prompts and data can stay on-prem. Product asks how quickly the team can tune it for a real workflow.

The winning model is often the one that survives all four questions.

Meta still has unmatched consumer-adjacent distribution and the strongest installed-base effect in the developer ecosystem. Mistral is easier to justify in many European enterprise settings. DeepSeek remains the reference point whenever the conversation turns to reasoning economics. Qwen may be the most industrially complete open family for multilingual and cloud-linked deployment. The winner changes depending on whether the buyer is a startup founder, a government procurement officer, a cloud architect, or a research engineer.

The result is a market that behaves less like one category and more like a layered stack.

At the bottom is availability: can I obtain the weights, compress them, and run them? In the middle is operability: do the license, docs, runtimes, and hardware requirements fit my use case? At the top is trust: when this system fails in production, who owns the problem?

Benchmarks matter. They just matter later than they used to.

The Real Fight Is Over Licenses, Data, and Geopolitics

The industry still likes the phrase “open-source AI” because it sounds morally clear and strategically attractive.

The market reality is messier.

Under the OSI definition published in 2024, open-source AI is not merely a model whose weights are accessible. Open-source models and open-source weights should include the data information and code used to derive those parameters. By that standard, the biggest names in the market are mostly not open source in the classic sense. They are open weight to varying degrees.

That distinction is no longer academic.

Open weights without data disclosure still create meaningful benefits. They enable self-hosting, fine-tuning, inspection, and cost control. They let governments and enterprises reduce dependence on closed APIs. They let smaller labs build on work they could never afford to reproduce from scratch. That is why the open-weight wave has been so productive.

But open weights also conceal where power remains concentrated.

Meta’s Llama models depend on training inputs that only Meta could assemble at that scale. Mistral’s Apache licenses are generous, but the company does not publish every layer of training and data curation. DeepSeek’s disclosures were detailed enough to transform the field, yet still incomplete enough to spark an entire reproduction effort. Qwen offers broad openness in weights and deployment but still sits inside a major cloud and platform company’s industrial logic. In all four cases, what is public is highly useful. What is private is still commercially decisive.

Geopolitics makes the issue harder.

Open-weight models have become instruments of sovereignty. In Europe, the attraction is obvious: downloadable models reduce dependence on U.S. API vendors and create room for local champions such as Mistral. In China and many emerging markets, open-weight systems from DeepSeek and Qwen offer strong capability, lower cost, and more local control than reliance on American closed models. Hugging Face’s January 2026 review argued that DeepSeek had been heavily adopted in Southeast Asia and Africa, precisely in markets where multilingual support, open-weight availability, and cost discipline mattered most.

At the same time, the same openness creates new anxieties. Western enterprises often want the economic benefit of open models without the geopolitical risk of relying on Chinese vendors. European regulators want innovation without ceding control over privacy or safety. U.S. companies want the ecosystem energy of openness without losing strategic control to their own users.

So license details matter more than they did a year ago.

Meta’s 700 million monthly-active-user clause is not just a quirky legal footnote. It is a statement about who gets to scale independently. Regional restrictions are not just compliance trivia. They determine where a supposedly open model can actually become infrastructure. Apache 2.0 is not only a developer convenience. It is a geopolitical selling point.

Safety complicates everything further. Open release accelerates innovation, but it also weakens any one vendor’s ability to control downstream behavior once the weights are widely copied and modified. Meta has tried to respond with detailed safeguard documentation and cyber-risk evaluations. DeepSeek’s R1 generated enough downstream activity that Hugging Face immediately moved to reconstruct and extend the training path. Mistral’s enterprise layer tries to restore trust through productization. None of these responses eliminate the basic trade-off. Open models lower barriers to good uses and bad uses at the same time.

The debate has matured too. The question is no longer whether openness is good or bad. The real question is simpler and harder: open which layers, for whom, under which obligations, and with whose data?

The companies winning today each answer that question differently.

Meta says: open the weights, keep the platform advantage. Mistral says: open the weights, monetize the service layer. DeepSeek says: open enough of the recipe to force the field forward. Qwen says: open at industrial scale and let distribution do the rest.

Those are not variations of the same strategy. They are competing political economies of AI.

The Center of Gravity Has Already Moved

Meta still matters. Llama did more than any other model family to make downloadable AI feel normal. But the market Meta helped create now rewards capabilities it does not own by itself: permissive licensing, cheaper reasoning, multilingual reach, rapid iteration, and easy deployment.

That is where Mistral, DeepSeek, and Qwen changed the picture. Mistral made openness legible to enterprise buyers. DeepSeek made reasoning portable. Qwen turned open release into an industrial process. None of them erased Meta’s advantage. All of them weakened its claim to be the only serious default.

The next phase will not be decided by one launch or one benchmark chart. It will be decided in ordinary rooms. Procurement reviews. Security audits. GPU budget meetings. Regional compliance checks.

One team will choose Apache 2.0 because legal can clear it in a day. Another will choose Llama because every tool already supports it. Another will reject a strong Chinese model because policy will not allow it. Another will choose Qwen because it already speaks the languages their customers use.

That is what the open-weight war looks like after the launch-day headlines fade. Not a philosophical debate. Not a leaderboard screenshot. A pile of operational decisions made by people who have to ship systems and live with the consequences.

The labs still write the model cards. More and more, the buyers write the market.

Published March 13, 2026. This analysis examines the competitive logic behind the open-weight AI market, with a focus on Meta Llama, Mistral, DeepSeek, and Qwen.

About the Author

Gene Dai is the co-founder of OpenJobs AI, focusing on AI-powered recruitment technology and the intersection of artificial intelligence with enterprise software.