Releases

Meta Superintelligence Lab released Muse Wonder and Muse Wonder Full, larger than Muse Spark and positioned as Meta's first frontier models. Meta says they score comparably to Fable 5.1 and GPT-6 Astra and lead on its MemoryBench continual-learning test (company claims). Muse Wonder costs $4 / $20 per million input/output tokens on the Meta Model API; Wonder Full is limited to Muse Code subscribers. Both power the Muse agent and were trained at the Prometheus data center in Ohio.

Meta released Muse Spark 1.1 and made it available through Meta Model API.

Meta announced Muse Video, a generative video model with native audio support, as coming soon.

Meta released Muse Image for instruction-following image generation, precise editing and multi-reference composition.

Meta released training code for Brain2Qwerty v1 and v2; v2 performs real-time non-invasive sentence decoding from brain recordings.

Meta Superintelligence Labs introduced Muse Spark, a multimodal model designed for efficient broad intelligence.

Meta released an efficiency update adding object multiplexing and up to 32-frame-per-second multi-object video tracking on one H100.

Meta released a tri-modal foundation model that predicts high-resolution human brain activity from video, audio and language.

Meta and World Resources Institute released a DINOv3-based global forest-canopy model with improved accuracy and detail.

Meta released a unified multimodal model for separating sounds from complex mixtures using natural prompts.

Meta released a model for estimating human body pose and shape from a single image.

Meta released a model for reconstructing 3D objects and scenes from a single image.

Meta released SAM 3 for text-, exemplar- and visual-prompted detection, segmentation and tracking in images and video.

Meta released speech-recognition models covering more than 1,600 languages plus an open corpus for 350 underserved languages.

Meta released DINOv3, a scaled self-supervised vision foundation model for high-quality visual features.

Meta released V-JEPA 2 for video understanding, physical-world prediction and zero-shot robot planning.

Meta released vision encoders for images and video designed to support recognition, localization and multimodal systems.

Meta released a multimodal safety classifier for Llama 4 inputs and outputs.

Meta previewed Behemoth as a teacher model for the Llama 4 family; it was not made generally available.

Meta released Llama 4 Maverick, a natively multimodal mixture-of-experts model for general AI workloads.

Meta released Llama 4 Scout, a natively multimodal mixture-of-experts model with a 10-million-token context window.

Meta released a behavioral foundation model for controlling virtual humanoid agents.

Meta released a multilingual 70B instruction model with performance comparable to larger Llama 3.1 models.

Meta published improved SAM 2.1 checkpoints and developer tooling.

Meta released a 90B multimodal model for image understanding.

Meta released an 11B multimodal model for image understanding.

Meta released a compact 1B text model for on-device use.

Meta released a compact 3B text model for on-device use.

Meta released SAM 2 for promptable object segmentation in both images and video.

Meta released a classifier for detecting prompt injection and jailbreak attempts.

Meta released a multilingual Llama 3.1-based safeguard model.

Meta released its 405B-parameter open-weights frontier model.

Meta released Llama 3.1 70B with a 128K context window and multilingual support.

Meta released Llama 3.1 8B with a 128K context window and multilingual support.

EvolutionaryScale, founded by former Meta researchers, released ESM-3, a generative protein model originating from Meta's ESM program.

Meta and World Resources Institute released global high-resolution canopy-height maps built with DINOv2.

Meta released the Llama 3-based update to its open safeguard model.

Meta released the 8B Llama 3 pretrained and instruction-tuned models.

Meta released the 70B Llama 3 pretrained and instruction-tuned models.

Meta released V-JEPA, a video model trained by predicting missing representations rather than pixels.

Meta released a 70-billion-parameter Code Llama model for code generation and instruction following.

Meta released Llama Guard, a safeguard model for classifying conversational AI inputs and outputs.

Meta released a low-latency streaming speech translation research model.

Meta released a research model that preserves vocal style and prosody in translated speech.

Meta released a more capable second version of SeamlessM4T.

Meta introduced Emu as the image-generation technology powering new generative AI experiences.

Meta released code-specialized Llama models for generation, completion and instruction following.

Meta released SeamlessM4T for multilingual speech-to-speech, speech-to-text, text-to-speech and text translation.

Meta released Llama 2 models in 7B, 13B and 70B sizes for research and commercial use.

Meta released MMS speech-recognition and synthesis models covering more than 1,100 languages.