Meta FAIR shares five research releases including SAM 2.1 and Spirit LM

Meta’s Fundamental AI Research (FAIR) team published a batch of open research releases in a single post, continuing its practice of releasing models and code rather than papers alone.

The headline releases were SAM 2.1, an updated version of Meta’s Segment Anything image- and video-segmentation model with better handling of visually similar and small objects, plus a training-code and demo package for developers; and Spirit LM, a multimodal model mixing text and speech that could generate more expressive, natural-sounding speech across dozens of languages. Alongside these, Meta released Layer Skip, an inference-acceleration technique that skips a subset of a model’s layers and verifies the result with the remaining ones, claimed to speed up generation by up to 1.7 times; SALSA, a machine-learning toolkit for stress-testing post-quantum lattice-based cryptography; and Meta Open Materials 2024, an open dataset and model pairing of roughly 100 million training examples for materials discovery, which the company said led its category on the Matbench-Discovery leaderboard.

None of the five was individually a frontier-model release; the significance was cumulative and strategic. It reinforced Meta’s positioning as the major lab most committed to open publication across a spread of domains — vision, speech, systems efficiency, cryptography and materials science — at a time when other frontier labs were disclosing progressively less about their methods.