The latest AI news and updates of Meta — the company behind Llama open-source models, FAIR research, and AI across Facebook, Instagram, and WhatsApp. Covering Meta's latest launches, research, and company news from the past 90 days.
MetaSep 8Meta launched Muse, a personal AI agent powered by its Muse Spark 1.3 model, designed to manage tasks within an isolated virtual machine. A host-side Sentinel agent governs all tool actions and network egress to ensure privacy. Meta also opened a bug bounty program offering up to $300,000 for valid security reports, including successful prompt injection attempts.
MetaSep 5Meta’s autonomous research system, AIRA3, placed 8th out of 4,000 teams in a live NVIDIA Kaggle competition, earning a Gold Medal. The system uses an ensemble of models and asynchronous agent coordination to autonomously improve model reasoning. Beyond the competition, AIRA3 has demonstrated capabilities in optimizing production GPU kernels and translating ancient Akkadian clay tablets.
MetaSep 2Meta released Muse Spark 1.3, featuring improved performance on long-horizon agentic and coding tasks. The model now actively collaborates by asking clarifying questions and confirming consequential actions, while reducing hallucination rates. Internal comparisons show a 20% reduction in tool calls and 25% fewer tokens. It is available today in Muse Code and Meta Model API.
MetaSep 1Meta Superintelligence Labs launched Muse Voice Transcribe, a real-time audio perception model for streaming ASR, diarization for over 20 speakers, and endpointing. The model uses reinforcement learning for adaptive delay, balancing speed and accuracy. It ranks first on Artificial Analysis benchmarks for streaming speech-to-text and is available via the Meta Model API, Meta AI for Mac, and Muse Code.
MetaAug 22Meta previewed WildArtifactBench, an internal framework for evaluating AI agents on complex, real-world tasks. Instead of fixed rubrics, it uses win rates and Elo scores from human and agentic preference judges to assess multimodal workflows. Meta released 10 tasks from the benchmark to help measure the practical utility that multimodal agents deliver in diverse deliverable formats.
MetaAug 22Meta released new evaluation results and demonstrations for Muse Spark 1.2, highlighting its visual reasoning and tool-use capabilities. The model now guides robots through unstructured environments, plans subtasks for bimanual manipulation, and generates captions for internal media training data. Muse Spark 1.2 is available via the Meta Model API and Muse Code.
MetaAug 10Meta released Muse Glimmer, a 30-billion-parameter open-weight model optimized for always-on local agent workflows. The model uses 4-bit quantization to fit under 20GB and a DFlash speculative decoding drafter to accelerate token generation on consumer hardware. It is available now under an Apache 2.0 license on Hugging Face, supporting multi-step agentic tasks and multimodal reasoning entirely on-device.
MetaAug 6Meta entered its AI models in five international STEM Olympiads, achieving gold-medal-level performance across all competitions without using external tools. The models secured perfect scores on the theory exams for both the Asian and International Physics Olympiads, alongside gold-medal results in the International Mathematical Olympiad, International Chemistry Olympiad, and the Romanian Masters of Mathematics.
MetaAug 5Meta released Muse Code (beta), a terminal coding agent powered by the new Muse Spark 1.2 model. The agent uses persistent async background agents and an append-only event log to manage long-horizon software engineering tasks with restart-safe execution. Muse Spark 1.2 features co-trained coding optimizations and is available today via the Meta Model API and terminal installation.
MetaJul 21Meta's SAM 3 and DINOv3 models are powering the Berkeley Lab-led SYNAPS-I project to automate scientific image segmentation. By pairing DINOv3’s global context with SAM 3’s pixel-level boundary extraction, the pipeline reconstructs 3D volumes in approximately 15 minutes. This replaces a manual labeling process that previously required a month of expert effort per dataset.
MetaJul 15Meta submitted an unnamed model to the Asian Physics Olympiad’s theoretical exam, where it achieved a perfect score of 30/30. The model tied for first place alongside the top three human student contestants. This result demonstrates the model's advanced reasoning and multimodal capabilities in solving complex, expert-level physics problems.
MetaJul 9Meta released Muse Spark 1.1, a multimodal reasoning model upgraded for agentic tasks including tool use, coding, and computer interaction. The model features multi-agent orchestration and a 1-million-token context window. The Meta Model API public preview provides developer access, while the Meta AI app and meta.ai offer the model in Thinking mode.
MetaJul 7Meta launched Muse Image, an agentic model that uses search and coding tools to refine image generation, and previewed Muse Video with native audio support. Muse Image is available in the Meta AI app, Instagram Stories, and WhatsApp. Third-party Arena benchmarks rank Muse Image second and Muse Video third in their respective categories.
MetaJun 29Meta released Brain2Qwerty v2, an end-to-end pipeline that decodes raw brain signals into text without surgery. The system achieves 61% average word accuracy, with top participants reaching 78%. Meta is releasing the training code for both v1 and v2, while its partner, the Basque Center on Cognition, Brain, and Language, is releasing the v1 dataset.
Frequently asked questions