AI at Meta
20 Intel
AI at Meta is Meta Platforms' official channel for artificial intelligence research and product news, run by the organization behind Meta's AI research division (originally founded in 2013 as Facebook AI Research, or FAIR). It shares updates on Meta's open-science AI work, including foundational contributions like PyTorch and the Llama model family. The account was formerly known as @metaai before rebranding to @AIatMeta.
Is this you? Sign in with X to claim this profile.
Intel
Meta AI released Muse Glimmer, a 30B-parameter open-weight model aimed at local, always-on agent workflows. It is designed to run on consumer hardware such as a Mac or a PC with a capable GPU, and ships under Apache 2.0. The announcement gives no benchmark figures.
Meta released Muse Code in beta, a terminal coding agent built on Muse Spark 1.2 that plans, implements and validates multi-file changes across large repositories using persistent sub-agents. Based on the announcement post; benchmarks and pricing are not covered here.
Meta Superintelligence Labs releases Muse Voice Transcribe, an autoregressive streaming ASR model. It reads 80ms audio chunks and decides per chunk whether to emit text, using RL for adaptive delay. It adds diarization for 20+ speakers and keyword and context biasing, and is available through the Me
Meta AI reports its models scored perfectly on the APhO and IPhO theory exams, took IMO gold, and reached gold-level on IChO and RMM, all with no search, coding or calculator tools. The post gives no model name or grading detail.
Meta describes a four-week research effort in which medical experts used Muse models to develop hypotheses, run analyses and draft ten manuscripts, with layered human review.
Meta details Muse Spark 1.2 multimodal capabilities: generating web pages and games from images or video with render-based self-correction, tool-assisted visual reasoning, and a robot orchestrator variant that decodes tool calls and loops until a task is done.
Meta previews WildArtifactBench, an internal framework that scores agents on real multimodal deliverables using win rates and Elo from human and agentic preference judges. Ten tasks are released. Only the announcement thread was read; the design-principles page was not available.
Meta's Muse Code runtime pairs a simple agent loop with persistent async background agents and an append-only local event log of every model call, tool run, approval and edit, making long-running sessions replay-exact and restart-safe. Based on the announcement thread only.
In long-horizon stress testing, Muse Code iteratively optimized GPU kernels across more than 1,000 tool calls and up to 24 hours on Hopper GPUs, with competitive gains over baseline Triton implementations for KDA and MLA.
Meta released six papers from mathematicians who used Muse Spark 1.1 and 1.2 in Thinking Mode, through the standard chat interface, on open problems. Humans guided the work, a second group reviewed it, and each paper marks human versus AI-drafted passages and credits prior work.
Meta's design team walks through Muse, an always-on personal agent with its own computer and browser. The post covers proactive messaging thresholds, human-in-the-loop gating of irreversible actions, goals tracking, and transparency features. The deeper technical post was not read.
Meta reports that its autonomous research system AIRA₃ took 8th of about 4,000 teams in a live NVIDIA Kaggle fine-tuning contest. Many isolated coding agents coordinate asynchronously through a shared forum and filesystem, with no central controller. Results are Meta's own and the thread is partial.
Meta released Muse Spark Open, an open-weight model with 816B total and 42B active parameters, a 1M token context window, and native text, image, video and audio input. Meta positions it for long-horizon coding and agentic work and calls it the most capable US open-weight model.
Meta introduces Muse Wonder and Muse Wonder Full, trained at its Prometheus data center. The models target personal and professional agents with Active Memory and Active Focus, and are available via the Meta Model API and Muse Code.
Meta released Muse Spark 1.3 in Muse Code and the Meta Model API, targeting longer-horizon agentic and coding work. It confirms before consequential actions and flags when stuck. Meta claims about 20% fewer tool calls and 25% fewer tokens than 1.2. Max reasoning and open weights are said to follow.
Meta releases weights for Muse Spark Open, a multimodal 816B-parameter MoE (42B active) with a 1M-token context, targeting coding and agentic work. Available via major API hosts and cloud catalogs, and runnable locally on two Apple M5 Ultra Mac Studios with MLX.
Meta released Muse Wonder and Muse Wonder Full, pitched at personal and professional agentic tasks with long-term memory. Wonder is on the Meta Model API and Muse Code, Full is limited to Muse Code subscribers, and the Muse app follows.
Meta released Muse Wonder and Muse Wonder Full, pitched at personal and professional agents with Active Memory and Active Focus. Available via the Meta Model API and Muse Code. Benchmark leadership claims are Meta's own and unverified.
Meta made max reasoning available for Muse Spark 1.3 in Muse Code and through the Meta Model API. The announcement is a short post, so benchmark numbers and pricing are not covered in what was collected.
Meta's technical report on Muse Glimmer, a 30B dense open-weight multimodal model distilled from Muse Spark for local agentic work. Covers architecture, why dense beat MoE under a 32GB limit, 4-bit quantization and speculative decoding, with self-reported agentic scores versus Gemma4-31B and Qwen3.6