Research models with a short summary of each: architectures, baselines and checkpoints.
A 7-billion-parameter base language model from Mistral AI that showed a small, well-trained model can compete with much larger ones. It is a popular starting point for fine-tuning. OrganizationMistral…
OLMo 2 publishes far more than weights: training code, intermediate checkpoints and the pretraining data mixes are open too. That makes it a strong choice for research into how language models learn. OrganizationAllen…
A compact model from Microsoft for settings where memory, compute or latency are tight. It reasons and does maths well for its size. OrganizationMicrosoftReleased2024Parameters~3.8BLicenseMIT Trained on filtered web and…
Give SAM a point or a box on an image and it returns a mask for that object, or let it mask everything automatically. It handles new image domains well without retraining. OrganizationMeta AI…
A family of vision transformer backbones trained without labels. Their features are good enough that a simple linear head often works for classification, depth or segmentation, with no fine-tuning. OrganizationMeta AI…
An encoder-decoder speech model that transcribes many languages, translates speech into English and detects the spoken language, all from one network. OrganizationOpenAIReleased2022Sizestiny (39M) to large (1.55B)…
One embedding model that does three kinds of retrieval: dense vectors, sparse lexical weights and multi-vector (ColBERT-style) matching. That makes it a good fit for hybrid search and RAG pipelines…