Here we share novel and promising architectures that may supplement or supplant other presently established models.
State-of-the-art generative models through iterative denoising
Embeddings compress a string of tokens into a high-dimensional representation. They are preferably contextually aware, meaning different strings of tokens...
The adversarial approach to generative modeling
Hybrid models combine multiple different architectures within a single system to reach a goal that no single architecture handles well on its own. Rather...
MOE provides the ability to use different smaller models that have better performance in certain domains. Their use is notable, as it has been stated that...
A multimodal model processes more than one kind of input, text, images, audio, video, within a single system, rather than requiring separate models stitched...
Reinforcement learning is a class of ML that uses dynamic feedback from an environment to reinforce successful outcomes.
The 2025-2026 shift toward small language models that run locally, the current model roster, and the runtime tooling behind it (Ollama, MLX, llama.cpp)
Transformers are a powerful type of architecture that allows input sequences to be considered with the whole input context. They are built on the...
Transformers that understand both images and text
The four distinct meanings of "world model" in 2026, from Dreamer-style latent planning to Sora/Veo video generation to Genie's interactive environments