# From Llama to World Models: 15 AI Models You Can Download in 2026 > Source: > Published: 2026-08-17 14:31:01+00:00 If most of your interaction with AI happens through ChatGPT, Claude or Gemini, it is easy to think of a model as something that lives behind a website or an API. There is another AI ecosystem growing alongside that one. The models come as files. You can put them on your own hardware, inspect them, fine-tune them, build products around them and, in some cases, see much more of how they were trained. What surprised me while looking through the current releases is how broad that ecosystem has become. A downloadable model in 2026 might be a language model like Llama. It might understand images and video. It might generate actions for a robot. It might predict how an object will move in 3D. NVIDIA’s latest Cosmos models go further and try to model possible states of the physical world. So I went through 15 models that are actually available to download today. This is not a ranking, and I am not trying to decide which one is “best.” I was more interested in what each release lets you do and how much of the model you really get. The phrase *open-source model* gets used rather loosely in AI. If a company releases the model weights, you can download the numerical parameters learned during training and run the finished model yourself. That is usually called an **open-weight model**. Open source can mean something more substantial. The [ Open Source Initiative’s definition](https://opensource.org/ai/open-source-ai-definition) includes the ability to use, study, modify and share the AI system, along with access to the model parameters, relevant code and sufficiently detailed information about the data and training process [1]. That difference matters because the models below do not all give you the same thing. Some provide weights and inference code. A few provide training code and datasets too. Others use custom licenses that deserve reading before you put the model inside a commercial product. With that out of the way, here is what is actually available. [ Llama 4](https://ai.meta.com/blog/llama-4-multimodal-intelligence/) is probably still the name most people associate with downloadable AI. Meta released Llama 4 Scout and Llama 4 Maverick as open-weight, natively multimodal models. Both use a Mixture-of-Experts architecture, which means only part of the model is active for a given input rather than every parameter being used every time [2]. Scout is the more accessible of the two. Meta says it can fit on a single H100 when quantized to Int4, while Maverick is substantially larger. Both can work with text and images, and the weights are available for developers to download. What Llama helped normalize is the idea that a capable general-purpose AI model does not necessarily have to remain behind somebody else’s API. You can host it yourself, adapt it and control the surrounding infrastructure. That does not mean Meta publishes everything required to reproduce Llama from scratch. Meta itself now uses the more accurate term *open-weight* for Llama 4. OpenAI returning to downloadable models was one of the more interesting changes in the open-model market. The [company released gpt-oss-20b and gpt-oss-120b](https://openai.com/index/introducing-gpt-oss/) under Apache 2.0. OpenAI says the smaller model can run within about 16 GB of memory, while the larger model fits within 80 GB [3]. That makes the 20B version particularly interesting. It is not just theoretically downloadable. It is within reach of relatively ordinary workstation hardware. OpenAI also provides reference implementations and allows developers to fine-tune and deploy the models on infrastructure they control. The company calls these *open-weight models*, not fully open-source reproductions. The original training corpus is not part of the download. For someone who wants an OpenAI model without sending every request to an OpenAI server, though, gpt-oss changes the options considerably. This is one of the freshest releases on the list. Alibaba’s Qwen team has now released the weights for [ Qwen3.8–2.4T-A95B](https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B), bringing a Qwen-Max-class model into its downloadable model family for the first time [4]. The numbers are enormous: 2.4 trillion total parameters with about 95 billion activated. The open checkpoint is designed for coding, research and long-running agent tasks, and it supports a native context length of 262,144 tokens that can be extended further. There is an important detail here. The downloadable Qwen3.8 checkpoint is not identical to the hosted Qwen3.8-Max service. The hosted version includes additional capabilities such as vision input, built-in tools and a larger default context configuration. The current Hugging Face release also uses a specific Qwen3.8-Max license rather than simply carrying the Apache 2.0 label familiar from earlier Qwen3 releases. This is a good example of why checking the actual model card is more useful than assuming every generation of a model family has the same terms. DeepSeek moves quickly enough that even an article written a few months ago is already dated. The official [ DeepSeek-V4-Pro-0813](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813) release arrived on August 13. Its weights are downloadable, and DeepSeek releases the repository and model weights under the MIT License [5]. V4 Pro is also a useful reminder that *downloadable* does not mean *easy to run*. The model has roughly 1.7 trillion parameters. You are allowed to run it yourself. Your laptop may have other opinions. That distinction matters. Open models remove dependence on a provider, but they do not remove the cost of memory, storage and computation. DeepSeek also publishes a smaller V4 Flash branch, which makes the family more interesting for people who care about deployment rather than simply having access to the largest checkpoint. [Gemma 4](https://ai.google.dev/gemma/docs/core/model_card_4) may be the most approachable family on this list if the goal is to experiment across very different hardware. Google released Gemma 4 under Apache 2.0 in several sizes, ranging from smaller E2B and E4B variants to 12B, 26B A4B and 31B models [6]. The smaller models target laptops and edge devices. The larger variants move toward consumer GPUs and workstations. Gemma 4 also accepts multiple input types, with support for text and images across the family and additional audio capabilities on some versions. That range matters more than it first appears. Open AI is often discussed as a competition between giant frontier models. In practice, a model that fits where you need to deploy it can be more useful than one that wins another benchmark but requires a rack of accelerators. Gemma 4 makes that tradeoff unusually visible. Mistral has consistently treated downloadable models as part of its product strategy rather than a side research project. [ Mistral Small 4](https://docs.mistral.ai/models) is a current Apache 2.0 model that combines instruction following, reasoning and coding in a single model [7]. There is something pleasantly uncomplicated about that. The weights are available under a familiar permissive license, and if you have the hardware, you can run and adapt the model yourself. Not every open-model release needs a complicated licensing story. For developers choosing between a hosted API and their own infrastructure, this kind of release makes the decision mostly an engineering question: capability, latency, privacy, hardware and cost. That is probably where open models become boring in the best possible way. Z.ai’s [ GLM-5.2](https://z.ai/blog/glm-5.2) is another major model whose weights are publicly available. The company released it under the MIT License and advertises a context window of up to one million tokens [8]. That combination makes GLM interesting for long-running coding and agentic workflows where the model needs to work with large amounts of material while remaining deployable outside a proprietary API. It is also another sign of how quickly the competitive map has widened. A few years ago, the downloadable-model conversation was dominated by a small number of labs. In 2026, serious releases are coming from Meta, OpenAI, Alibaba, DeepSeek, Google, Mistral, Z.ai, Ai2, NVIDIA and others. There is no longer one obvious open-model ecosystem. There are several overlapping ones, with different licenses, hardware requirements and communities forming around them. [ Kimi K2.5](https://github.com/MoonshotAI/Kimi-K2.5) is where the line between a language model and an agent starts becoming less clear. Moonshot AI describes K2.5 as a native multimodal agentic model. It has one trillion total parameters, with 32 billion activated, and was built through continued pretraining on a large mixture of text and visual data [9]. It can work with images as well as text and is designed heavily around coding and tool use. The weights and repository are available, but the license deserves more attention than a casual glance at the word “MIT.” Kimi uses a modified license with additional terms, including an attribution requirement for very large commercial products. That is not necessarily a problem. It is simply something developers should know before treating every downloadable model as legally interchangeable. The model file is only one part of the decision. If you want to see what **fully open AI research** looks like, OLMo is the model I would examine. [ Ai2’s OLMo 3](https://allenai.org/papers/olmo3) family includes 7B and 32B models, but the parameter count is not the most interesting part. Ai2 releases the entire model flow: data, code, checkpoints and training details [10]. That changes what a researcher can do. With a normal open-weight model, you can study the finished system. With OLMo, you can look much further upstream by inspecting how the data was assembled, examining intermediate checkpoints and studying how the model changed during training. For an everyday user, that level of openness may not matter much. For researchers trying to understand why a model behaves the way it does, it is a significant difference. [ Molmo 2-O](https://allenai.org/blog/molmo2) takes the same philosophy into multimodal AI. The 7B model is built on OLMo 3 and can work with images and video. The broader Molmo 2 release includes open weights, training data and training code [11]. What makes Molmo particularly interesting is grounding. A multimodal model should not only recognize that a person or vehicle appears in an image. Molmo can also point to objects, track them and reason about visual information across frames. That becomes important once AI starts interacting with software, robots or physical environments. It needs to connect language to specific things in space and time. Describing what is visible is useful; knowing where something is, and how it changes across frames, starts to make that perception actionable. The first ten models still look familiar if your mental picture of AI is a chatbot. The final five are where the list moves beyond the usual chatbot model. They are examples of the model itself moving closer to perception, prediction and action. [ Physical Intelligence’s π0](https://www.physicalintelligence.company/blog/openpi), pronounced pi-zero, is a generalist robot policy. Instead of predicting the next word, it predicts actions for a robot. The openpi release provides model weights, code, several fine-tuned checkpoints and code for adapting π0 to other tasks and robot platforms [12]. That changes the meaning of “running a model.” You do not type a question and wait for a paragraph. The model receives observations and instructions and helps determine what a physical machine should do next. This also exposes why open robotics is harder than open language modeling. Having the weights is useful, but you also need the right sensors, action representation, robot hardware, training data and safety controls. This also shows why open robotics is harder than open language modeling. Having the weights is useful, but you still need the right sensors, action representation, robot hardware, training data and safety controls. In other words, you can download the brain. The body is still your problem. [ Cosmos 3](https://github.com/NVIDIA/cosmos) is the model family that makes the phrase NVIDIA describes Cosmos 3 as a family of omnimodal world models for physical AI. The models can work across text, images, video, ambient sound and action sequences, and are intended for tasks including world generation, physical reasoning and action prediction [13]. In simpler terms, a normal image model might look at a scene and describe what is there. A world model is interested in what could happen next. If a vehicle moves forward, how might the scene change? If a robot takes an action, what state might follow? If we change the weather, lighting or motion in a synthetic environment, does the resulting future remain physically plausible? Cosmos 3 models and code are available to download, including different sizes and specialized checkpoints. This is where open AI starts to look much less like an alternative chatbot market and much more like infrastructure for robotics, autonomous systems and simulation. [ GR00T N1.7](https://developer.nvidia.com/isaac/gr00t) takes another step toward action. It is a **vision-language-action model**, usually shortened to VLA, built for robotics. NVIDIA makes the model weights available and provides the codebase for developing and adapting GR00T policies [14]. A VLA model connects three things that were traditionally treated more separately: visual perception, language understanding and physical action. That makes these models fundamentally different from systems whose job ends when they generate text. It also means mistakes have a different meaning. A bad chatbot answer is annoying. A bad action sent to a physical machine can have consequences. Downloading the model is therefore only the beginning. Simulation, action limits, evaluation and human oversight matter much more once the output can move hardware. Ai2’s [ MolmoAct 2](https://allenai.org/blog/molmoact2) is another open vision-language-action model, but its release is particularly interesting from a research perspective. MolmoAct 2 builds on an embodied version of Molmo and adds an action component that generates continuous robot controls. Ai2 has also released checkpoints for further robot fine-tuning rather than presenting the model as a universal plug-and-play controller [15]. That distinction is important. You cannot download a robot foundation model, connect it to an arbitrary arm and expect useful behavior five minutes later. The interesting part is that the ingredients needed to experiment with that problem are becoming accessible. Ai2 has published the model, training material and supporting components much more openly than is typical in robotics. That gives researchers a way to inspect not just the final behavior but more of how it was produced. The last model on my list does something I find especially interesting. [ MolmoMotion](https://allenai.org/blog/molmo-motion) does not mainly describe what is happening in an image. It tries to predict how something will move next. Given visual observations, points on an object and a written action such as “move and rotate the bowl,” the 4B model predicts future 3D trajectories for those points [16]. Ai2 also released a large motion dataset and a benchmark alongside it. This may sound narrower than a general chatbot, but the underlying problem is fundamental. Perception tells us what has already happened; intelligent action also requires some idea of what is likely to happen next. A robot reaching for an object, an autonomous vehicle planning around another road user and a world model generating a plausible future all depend on some version of that question. **Where will the world be a moment from now?** That is a very different AI problem from predicting the next word. Going through these releases changed the way I think about the open-model race. Most comparisons focus on capability: which model reasons better, codes better or supports a larger context window. Those questions matter, but another change is happening at the same time. The range of things we can download is expanding. Llama provides a multimodal general-purpose model. gpt-oss puts an OpenAI reasoning model on hardware you control. OLMo exposes much more of the training process, while Molmo connects language with visual grounding. π0 and GR00T turn model outputs into robot actions, MolmoMotion predicts physical movement, and Cosmos tries to represent possible futures of an environment. These are not fifteen versions of the same product. They are beginning to occupy very different parts of the AI stack. “You can download it” is not the same as “you can run it tonight on a MacBook.” Some models on this list are genuinely accessible on consumer hardware, while others require high-end GPUs or much larger infrastructure. DeepSeek V4 Pro and the largest Qwen releases, for example, remain substantial compute projects even though their weights are available. I find it useful to separate two questions. First, do I have permission to obtain, modify and deploy the model? Second, do I have the hardware and engineering resources to do it practically? A permissive license can solve the first problem. It does not manufacture VRAM. If you have a normal workstation and simply want to experiment locally, **gpt-oss-20b** or one of the smaller **Gemma 4** models are sensible places to start. They are much closer to the kind of model an individual developer can realistically run without building a small data center first. If you care more about research and reproducibility, I would look at **OLMo 3** and **Molmo 2-O**. Ai2 releases much more than the final weights, which makes these models especially useful if you want to understand how the system was built rather than only use it. For robotics, **π0**, **GR00T N1.7** and **MolmoAct 2** are three useful entry points with different strengths. And if world models are the part of this list that caught your attention, **Cosmos 3** is probably the most obvious place to start exploring. I would not begin with the largest model just because it has the largest number next to its name. For local AI, the model you can actually run, modify and learn from is usually the more useful one. I do not think downloadable models will eliminate closed AI services. Most people do not want to operate a model any more than they want to maintain their own email server. Hosted systems are convenient, and proprietary frontier models will continue to have advantages. Open weights do change what can happen underneath an application. A company can keep sensitive data on its own infrastructure. A researcher can inspect and alter a model instead of only sending it prompts. Smaller models can live on edge devices, robotics teams can adapt VLAs to their own hardware, and world models can be post-trained around particular environments. That control becomes more important as AI moves from generating text toward observing, predicting and acting. A few years ago, downloading an AI model mostly meant getting something you could talk to. In 2026, the same idea now covers models that see, point, track, reason, predict motion, control robots and generate possible worlds. That is a much more interesting shift than who happens to have the best chatbot this month. [1] Open Source Initiative, “The Open Source AI Definition 1.0.”[https://opensource.org/ai/open-source-ai-definition](https://opensource.org/ai/open-source-ai-definition) [2] Meta, “The Llama 4 herd: The beginning of a new era of natively multimodal AI innovation,” April 2025.[https://ai.meta.com/blog/llama-4-multimodal-intelligence/](https://ai.meta.com/blog/llama-4-multimodal-intelligence/) [3] OpenAI, “Introducing gpt-oss,” August 2025.[https://openai.com/index/introducing-gpt-oss/](https://openai.com/index/introducing-gpt-oss/) [4] Qwen, “Qwen3.8–2.4T-A95B,” official model release, August 2026.[https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B](https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B) [5] DeepSeek, “DeepSeek-V4-Pro-0813,” official model release, August 2026.[https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813) [6] Google DeepMind, “Gemma 4 Model Card,” 2026.[https://ai.google.dev/gemma/docs/core/model_card_4](https://ai.google.dev/gemma/docs/core/model_card_4) [7] Mistral AI, “Models Overview: Mistral Small 4,” 2026.[https://docs.mistral.ai/models](https://docs.mistral.ai/models) [8] Z.ai, “GLM-5.2: Built for Long-Horizon Tasks,” June 2026.[https://z.ai/blog/glm-5.2](https://z.ai/blog/glm-5.2) [9] Moonshot AI, “Kimi K2.5,” official repository.[https://github.com/MoonshotAI/Kimi-K2.5](https://github.com/MoonshotAI/Kimi-K2.5) [10] Ai2, “OLMo 3,” 2025–2026.[https://allenai.org/papers/olmo3](https://allenai.org/papers/olmo3) [11] Ai2, “Molmo 2,” December 2025.[https://allenai.org/blog/molmo2](https://allenai.org/blog/molmo2) [12] Physical Intelligence, “Open Sourcing π0,” February 2025.[https://www.physicalintelligence.company/blog/openpi](https://www.physicalintelligence.company/blog/openpi) [13] NVIDIA, “Cosmos 3,” official repository and model family, 2026.[https://github.com/NVIDIA/cosmos](https://github.com/NVIDIA/cosmos) [14] NVIDIA, “Isaac GR00T,” official developer platform.[https://developer.nvidia.com/isaac/gr00t](https://developer.nvidia.com/isaac/gr00t) [15] Ai2, “MolmoAct 2,” May 2026.[https://allenai.org/blog/molmoact2](https://allenai.org/blog/molmoact2) [16] Ai2, “MolmoMotion: Language-guided 3D motion forecasting,” June 2026.[https://allenai.org/blog/molmo-motion](https://allenai.org/blog/molmo-motion) [From Llama to World Models: 15 AI Models You Can Download in 2026](https://pub.towardsai.net/from-llama-to-world-models-15-ai-models-you-can-download-in-2026-d88aba951dec) was originally published in [Towards AI](https://pub.towardsai.net) on Medium, where people are continuing the conversation by highlighting and responding to this story.