Open source AI icons depict accessibility, collaboration, and community support. getty
Open source AI is usually framed as the scrappy option. You reach for it when budgets are thin, procurement is slow or your team wants to experiment without handing sensitive work to a remote API. That view now feels behind the market. Open models have become materially better, the software for serving them has grown up, local runtimes have become ordinary developer tools and the agent layer is now spilling into public repos at a pace the large commercial labs can no longer shrug off. The question is no longer whether open source AI matters. The real question is whether companies can still afford to treat it as secondary, and if the large vendors can continue to ignore it.
Open source AI is rising fast, but the gains do not show up evenly across the market. Menlo Ventures’ 2025 enterprise report says open source LLMs account for only 11% of enterprise market share by production API usage, down from 19% a year earlier. Large companies still lean toward closed vendors when they want managed service, support and less operational burden. Yet the same report says developer interest has spread across Chinese model families and platforms such as Qwen from Alibaba, DeepSeek’s V3 and R1 series, Kimi from Moonshot AI, MiniMax’s model lineup, and GLM from Z.ai. That split might reflect the fact that enterprises often move slowly while developers usually are more eager to test the latest models.
The public ecosystem numbers show a much stronger trend. Hugging Face, a public hub for open models, datasets and AI tooling said in its Spring 2026 review that its network had grown to 13 million users, with more than 2 million public models and more than 500,000 public datasets. It also said more than 30% of the Fortune 500 now have verified accounts on the platform. This shows that a public market for open AI components is taking shape.
Open source models, especially the Chinese ones, are starting to perform at levels that match some of the most popular models from OpenAI, Anthropic and Google. Stanford’s 2026 AI Index says the performance gap between the top closed model and the top open model continues to narrow. As of March 2026, the best closed model led the best open model by 3.3% and six of the top ten models on the Arena Leaderboard were closed. So while closed AI models from commercial vendors still lead the pack, the performance floor for what these open source models can do is getting much higher and the stack around open models getting good enough to competitively perform real work.
Not too long ago, the top of the line models were GPT 5.3 or Claude Opus 4.5, and now open source models perform as well or better than both of those models, even as Anthropic and OpenAI continue to iterate and improve.
The open source story impacts across several layers at once. The first layer is the models themselves. The most direct line for open source is to reduce “open weight” models. An open weight model is an AI model whose trained internals are released so others can run it themselves, rather than only renting access through an API.
Meta’s Llama line was one of the first of the major models to be released with open weights. Soon, open weight models from Mistral, Gemma, Qwen, DeepSeek and Moonshot widened the field. OpenAI even made a decision to release gpt-oss-120b and gpt-oss-20b under Apache 2.0 open source license in August 2025. These open weight OpenAI models are capable models that can run on local machines disconnected from the Internet and OpenAI’s data centers. The important nuance is that open weight does not always mean fully open source. A model can have open weights but still keep other pieces closed, such as the training data, the training code, the fine-tuning process and have non-open license terms.
The second layer is inference and serving infrastructure. A model is not enough on its own. It is like having a powerful car engine sitting on the floor of a garage. It might be impressive, but it is not useful until someone builds the rest of the car that lets it run smoothly, cheaply and at scale. That is what inference and serving infrastructure does. It is the software that helps companies actually run AI models in the real world. It decides how fast the model responds, how many people can use it at once, and how much hardware and memory it burns through. Open source infrastructure vLLM has become one of the most important open projects in the space, describing itself as a high throughput, memory efficient serving engine supported by more than 2,000 contributors.
The third layer is local runtime software. This layer is about running AI on your own machine instead of always calling someone else’s service over the internet. Two years ago, running a model on your own hardware was still a fiddly, technical project for specialists. Tools like Ollama, LM Studio, and llama.cpp have made local AI far easier to use, which means a developer or company can download a model and run it locally on a laptop, workstation or internal server, and also make it available for others to use on the same machine or over the network.
The fourth layer is agents, and this may be where the open ecosystem is moving fastest. LangChain’s 2026 State of Agent Engineering survey found that 57% of respondents already had agents in production. The shift is that open source agents are starting to look less like chatbot add-ons and more like the bones of a full worker stack. OpenClaw, a viral agentic AI solution that burst onto the scene a few months ago, gives users an open, self-hosted assistant layer that can sit on hardware they control and plug into real communications and task channels. Another open source initiative gaining steam is the Hermes Agent, which pushes the idea further toward long-lived autonomy, with memory, reusable skills, and a learning loop that developer Nous Research says can run on anything from a cheap VPS to a larger cluster.
Competing with OpenAI’s Codex and Anthropic’s Claude Code are additional open source initiatives. OpenCode brings the same open logic into software development, offering an open source coding agent that is explicitly provider-agnostic and can work with closed APIs or local models. Once those pieces are combined with open-weight models and open-serving software, the result starts to look like a real alternative stack, not a hobbyist patchwork. The open market is steadily assembling its own answer to the tightly integrated AI platforms now being built by Anthropic, OpenAI, Google, and Microsoft.
While open source offers an increasingly credible alternative stack, it is not yet a clean replacement for every closed platform. The commercial vendors still have stronger polish, managed reliability, and tighter product integration in many cases. They also package the best performance, polished safety layers, managed uptime and a single procurement and buying experience. Yet the closed model market has developed cracks that nudge customers toward open alternatives.
The first is cost. Usage pricing is easy to tolerate in experimentation but much harder to ignore in steady production. Teams serving internal copilots, research workflows or agent systems can watch a small test turn into a real operating expense with surprising speed. OpenAI and Anthropic’s continuing evolution of their pricing make the point as they increasingly meter usage of flagship models, long outputs and tool use.
At the same time, subscription plans across the market are starting to feel more metered than they first appear. Microsoft now spells out AI credits and feature limits in consumer Copilot plans. Google’s AI Ultra Access includes a monthly credit allotment and even supports paid overages beyond that cap. Anthropic’s recent OpenClaw-related restriction policy change pushed some power users toward separate pay-as-you-go billing. The more AI behaves like always-on infrastructure, the less comfortable buyers become with paying for it as if it were a series of casual chats and the more they hit what had previously been seen as very generous subscription limits.
This does not mean open source is always cheaper in absolute terms. Self-hosting brings hardware, engineering and operations costs of its own. However, for sustained, high-volume or autonomous workloads, many buyers are deciding they would rather own more of the cost structure than stay fully exposed to token billing, subscription usage caps and the need to pay for overage usage.
Another issue is control. Companies handling regulated data, internal code, legal material or sensitive customer records often want a clearer grip on where prompts run and where outputs end up. This pushes towards full control of the AI stack that answers the questions that boards and compliance teams are increasingly asking: Where does the system run? Who can inspect it? Can we move it if a vendor changes terms?
A third pressure is platform dependence. Once an organization builds around one vendor’s APIs, they become increasingly dependent on rate limits, changes to per-token costs and changes to model behavior as the vendors upgrade or deprecate models. For many non-U.S. players, open source AI is starting to look like a sovereignty strategy. The European Parliament warned in 2025 that Europe’s dependence on foreign technology weakens its room to act and leaves sensitive data exposed to outside legal and political pressure. In January 2026, the European Commission opened work on an Open Digital Ecosystem Strategy tied directly to technological sovereignty. Those moves suggest that open AI is being viewed not just as a developer preference, but as a way to avoid the latest evolution of vendor lock-in, this time at the model and agent layer.
China’s role in open source AI is one of the main reasons the conversation around open source AI has changed so quickly. Hugging Face said in March that China had surpassed the United States in both monthly and overall downloads on its platform and that Chinese models accounted for 41% of downloads over the past year. It also said that most trending models released in 2025 were either developed in China or derived from Chinese models.
Chinese labs are not just trying to offer cheaper alternatives to Western reasoning models, they are building open systems suited to real work across software, automation and multimodal tasks. OpenClaw and Hermes users now treat China-based Moonshot’s Kimi K2.5 and Z.ai’s release this month of its GLM-5.1 next-generation flagship models as serious open model options for agentic use.
The biggest commercial vendors still hold the premium end of the market. They still lead plenty of benchmark categories, and their adoption and revenue show no signs of decreasing. Yet several moves suggest they do not view open source as just background noise.
The clearest recent example was DeepSeek. When the Chinese lab released its open-weight R1 reasoning model in January 2025, it helped trigger a broad AI stock selloff, wiping roughly $593 billion off Nvidia’s market value in one day and forcing investors to rethink the cost structure behind the entire boom. OpenAI released its open weight models in response. Now with OpenClaw’s surge in popularity, Anthropic is forced to respond to the open source offering by drawing clearer boundaries around how third-party agent tools consume Claude. Even Microsoft is reportedly testing Copilot capabilities inspired by OpenClaw-like autonomous task execution.
There is another reason for commercial labs to pay attention. Open source does not need to beat them everywhere. It only needs to become good enough, cheap enough and flexible enough for a growing slice of important work. Once that happens, price pressure rises and customers begin to explore their options to move between hosted and self-run setups.
So is now the time for open source AI?
For many teams, yes. Not in the sense that every company should throw out commercial vendors tomorrow, and not in the sense that open source has already taken the top of the leaderboard and locked it down. The more realistic sense is that open source AI now has credible models, mature infrastructure, workable local runtimes and a live agent ecosystem that is moving quickly enough to change buyer behavior. China’s model labs have added speed and competitive pressure to that shift.
Projects like OpenClaw and Hermes show that the agent layer is not going to stay inside closed products. Moonshot’s Kimi models and GLM models from Z.ai show that Chinese open model providers are now part of the same stack implemented in real-world, high importance projects. The old way to think about open source AI was as the backup plan. The market in 2026 does not really support that view anymore. Open source is now part of the main route.
This article was originally published on Forbes.com