Meta
An update to SAM 3, carrying the same restrictive license.
A variant of SAM to segment audio from videos.
An update to the well-known image segmentation model.
A code generation model trained on Docker-agent trajectories and Python execution traces, released under a non-commercial license.
A small LM released under a non-commercial research license.
A new version of Dino, the foundation vision models by META. However, they did not reuse the Apache 2 license from the previous versions but went with a custom one which works similarly to the Llama License.
Kernel-writing models and methods, which either generate triton or CUDA kernels, are becoming increasingly popular. KernelLLM is a relatively small model, based on Llama 3.1 8B, capable of converting PyTorch code to triton kernels.
A foundation model for molecular chemistry. The dataset is also open-sourced.
After our coverage of the initial release, LMArena (which now formed a company and has a new look, accessible here) announced they will update their policies and released a comparison HF space. They also brought the regular instruct version of Llama 4 Maverick into the Arena. While the experimental version from META scored an Elo of 1417, the instruct version scored only 1271, landing it on place 35 (24 with style control). However, this should not be seen as yet another dunk on Llama 4 - we found the regular version more pleasant to use compared to the experimental version deployed by META. The chat version (over)uses emojis and likes to please the user by saying its question is interesting. The comparison HF space is really interesting, as it also gives a glimpse into the questions (LMArena) users care about.
Meta was very vague in their announcement for this model, but our take (and backroom conversations) is them updating a bunch of their post-training stack from just rejection sampling and DPO to more in-vogue RL methods (like our Tulu).
A series of small language models by Meta, ranging from 125M to 1B parameters.
A strong LLM-as-a-judge model trained by bootstrapping all of its own data! Really cool research.