Open models always come back

##Introduction

Free software has a curious habit of always being on the verge of taking over or disappearing. We have spent decades waiting for the year of Linux on the desktop, then burying it again a few months later. Something similar is starting to happen with AI models: open models also come back in waves.

Over the past few weeks, models with downloadable weights have appeared across almost every modality. Kimi K3 has moved closer to the best proprietary language models. Ideogram and Krea have released new image models. MiniMax has opened part of H3, while Lightricks has launched LTX-2.5 through its API alongside a collection of downloadable weights. Black Forest Labs has also shown its first video model and already announced a future open version of FLUX 3.

It is tempting to put all these releases together and declare that proprietary models are losing their lead. The problem is that many of these models are not really open source. Some do not allow commercial use. Others keep their best version closed, and H3 excludes several of the world's largest markets.

That does not make the wave an empty marketing campaign. It makes it quite a bit more interesting. In this article we will see why releasing weights is also a way to compete, what these companies are trying to achieve, and how much freedom users actually keep.

##Open, but what exactly?

A model is not a single thing. There are the weights produced by training, the code required to run them, the data, the training process itself, the license, and everything that turns the model into a product. Releasing one part does not automatically open the rest.

When a company publishes weights that we can download and run, we call them open weights. If the license also lets us use, modify, and redistribute them for any purpose, it grants us real freedom. To call the model open source in a stricter sense, we would also need the training code and enough information about the data and process to study and reproduce the system. Broadly speaking, that is the distinction made by the OSI's Open Source AI Definition.

Almost all these models also rely on open libraries such as PyTorch, Transformers, Diffusers, vLLM, or SGLang. Training and running them would be much harder without those libraries, but using open-source software does not automatically make the model open source. A closed model can run entirely on open tools.

That is why two downloadable models can provide completely different freedoms:

Z-Image, from Tongyi-MAI, is the simplest case. Alibaba releases a 6B base image model, the code required to run it, and an Apache 2.0 license. It does not release the data or the complete training process, so open weights remains the more precise term, but we can modify the model and use it commercially without asking for another license.

Krea, Ideogram, and MiniMax also publish weights that we can download and run. The difference lies in what their licenses allow and which parts each company keeps closed. The weights answer whether we can run the model. The license determines what we are allowed to do with it.

##A wave across every modality

A single open-weight release may be an experiment. When several arrive at roughly the same time across such different fields, it starts to look like a trend.

##Why they always come back

This is not the first time open models have looked ready to catch up with proprietary ones. It will not be the last either.

Stable Diffusion released its weights in 2022 and made it possible to generate and modify images at home. The community built interfaces, LoRAs, ControlNets, training systems, etc. No single provider could have developed all of that alone. Midjourney had already been popular for some time when V6 widened the gap again at the end of 2023. FLUX.1 [dev] narrowed part of it in 2024, then Nano Banana moved the bar again the following year.

Stability AI also shows that massive adoption does not guarantee a sustainable business. In 2024, after exceeding 150 million downloads, the company changed direction and raised another round of funding. Much of the value created by Stable Diffusion ended up in interfaces and providers that Stability AI did not control. We cannot blame openness for the company's problems, but getting adoption and knowing how to monetize it are different problems.

Something similar happened with language models. GPT-4 opened a clear gap in 2023. Llama 3.1 moved closer again in 2024, then the updated Claude 3.5 Sonnet took the lead in coding and tool use. DeepSeek-R1 narrowed the gap and Claude Opus 4 widened it again during 2025.

In video, the cycle began later and moved faster. Kling raised the quality of proprietary models in 2024. Wan 2.1 released weights and code a few months later. Veo 3 widened the gap again with native audio, while MiniMax H3 and LTX-2.5 have just narrowed it with downloadable weights. Each modality repeats the same movement at its own pace.

The AI Index 2026 shows exactly this back and forth. The gap between the best open and closed models narrowed until it almost disappeared in 2024, then widened again during 2025. The exact figure depends on the benchmark, but the pattern remains: the gap opens and closes.

Being the first to develop a new capability remains extremely expensive. Once someone has shown that it works, the cost of reproducing it falls much faster thanks to algorithmic progress and hardware. Epoch AI estimates that capabilities from the proprietary frontier have historically reached downloadable models several months later.

Open models do not need to remain in first place. It is enough for them to turn an exclusive advantage into a foundation that others can build on. Stable Diffusion did not have to remain the best image model to change the market. It only had to enable things that no API allowed at the time.

##Releasing weights is also a way to compete

Releasing weights costs money. A company has to prepare checkpoints, write inference code, document the model, and support a community that will find quite creative ways to break it. If so many companies do it, it is not because they suddenly stopped wanting to build a business.

The first reward is distribution beyond the company's own API. A downloadable model quickly appears in ComfyUI and Diffusers, as quantized versions, and through inference providers. Each integration makes the model cheaper to try and more likely to become the foundation of somebody else's work. The company is no longer the only one able to adapt it to new uses.

For Alibaba, the connection to its business is direct. The company has explained that more open models lead to more applications, and those applications eventually consume cloud resources. Releasing the model can be a way to sell the infrastructure that runs it.

Wan and Qwen suggest that this strategy is changing. Alibaba still keeps important models open, but the latest versions of several families have remained in its API. Openness may have already delivered the distribution the company wanted, and now it may want to monetize a larger share of the usage. The choice could also be temporary or depend on the cost of each model. Alibaba has not explained the reason, so for now the pattern matters more than our theory.

Krea and Ideogram follow another strategy. Both release competitive models, but reserve their most capable versions and some parts of the product for their APIs. Krea 2 Raw and Turbo can be downloaded, while Medium and Large live in the API. Ideogram lets us experiment with FP8 and NF4, but requires a commercial license to use them in production.

This also lets them charge according to the size of the business using the model. The Krea 2 license grants commercial use to companies with less than $1 million in annual revenue. Above that threshold, they have to negotiate an Enterprise license. A developer can build with the model, while Krea keeps the option to charge once that use grows into a sizable business.

Openness can also be a commercial strategy. It can lead users to the API, the cloud, or a higher-tier license. In exchange, the company accepts that others can run the model and build products it does not control.

##When open weights become a marketing tool

Calling a model open has also become part of how companies market it. The problem begins when the announcement suggests more freedom than the license or the product actually provides.

##The creator does not have to be the operator

I work at Runware, an inference provider. For us, having the weights means we can run the model, optimize it, and add features without depending on the original lab to offer them.

That does not mean everyone should buy several GPUs and deploy Kimi K3 at home. Its documentation recommends dozens of accelerators. Even much smaller models require updates and monitoring, then enough steady usage to amortize the hardware. In many cases, using a provider remains cheaper and quite a bit simpler.

The difference is that the creator stops being the only operator. Different providers can compete on price, latency, availability, features, etc. If none of them works for you and you have the necessary infrastructure, you can run the model yourself.

That ability to switch providers or run the model myself is the freedom I care about most. I do not need every user to self-host. I need them to be able to choose who runs the model and to have an alternative if the API changes its pricing, disappears, or stops allowing their use case.

Open weights do not eliminate intermediaries. They prevent any one intermediary from becoming mandatory.

##Why this wave comes from China

Many of the models driving this wave come from China: Kimi, DeepSeek, Qwen, GLM, Wan, H3, Z-Image, etc. This does not look like a coincidence, but there is no single explanation either.

Competition within China is intense. Releasing weights allows a model to gain developer adoption and integrations without having the distribution channel of ChatGPT or Gemini. It also encourages derivative models and lets companies in other countries offer them without forcing their users to depend on a Chinese API.

For companies selling cloud infrastructure, hardware, or services, releasing the model can attract users to the rest of the business. Alibaba does not always need to charge for the checkpoint if its ecosystem ultimately generates revenue elsewhere on the platform. Openness is not a donation. It is a way to choose which part of the chain the company wants to compete in.

Chinese government policy also explicitly supports open ecosystems. On top of that, companies face pressure to achieve more with less access to the most advanced hardware. It is reasonable to think that chip restrictions encourage efficient architectures and optimization across very different hardware. What we cannot claim is that those restrictions directly cause companies to release their weights.

The result can be measured. Hugging Face estimates that Chinese models account for around 41% of open-model downloads on its platform. A download is not the same as a production deployment, but this is more than a collection of announcements.

There is something ironic about all of this. For years, open software was presented as a way to avoid depending on large technology companies. Now large Chinese companies are using it to compete with other large companies. It is not exactly the revolution that the early free software advocates imagined, but the user still gets more choices.

##They do not need to win to change the market

Proprietary models still have advantages. They tend to introduce the hardest capabilities first and can provide a more polished experience. They also do not require the user to maintain infrastructure. In some modalities they remain clearly ahead. Denying that would only weaken the argument.

Open models are not always cheap or transparent either. And being able to download a checkpoint does not mean it is easy to run. It may occupy hundreds of gigabytes and require a lot of hardware. Even then, it may depend on a part that the company has not released.

Even so, every wave means capabilities spend less time locked away. What existed only inside an API yesterday becomes available through other providers, and the community can adapt it in ways the creator would never have built.

I prefer that world. Not because I expect open models to lead every benchmark, but because they force companies to compete beyond the model itself. Training it is no longer enough. Who can run it and how much freedom the user keeps also matter.

Will 2026 be the year of open models? I do not know. It does not need to be. As long as they keep giving users back the ability to choose, this wave has already accomplished something.