AI Lab
Official Meta AI announcements — model releases, product launches and research, each one summarized with every source covering it, by AIBriefs. RSS
Catching JiTTests, which use LLMs to generate executable tests proving a behavioral change before code lands, is now part of Meta's infrastructure rather than a research prototype. Meta says the approach catches bug classes that LLM- or agent-based code review misses, with less load and fewer false positives.
Meta's next-gen Andromeda ads retrieval engine hits over one billion ads per second, a 100x lift over the 2024 system, via model-system-GPU co-design. It has delivered a 2.4% increase in conversions on Facebook since 2025.
Meta replaced per-item IDs with a stable semantic token vocabulary, letting its Generative Ads Model reason about unseen products instead of starting from zero signal. Paired with GEM and Meta's Adaptive Ranking Model, it delivered a +0.35% conversion lift across surfaces for advertisers.
Meta built a 1,300-line orchestrator that does zero domain work itself, routing 20+ data-pipeline workflows to 14 specialized agents. Its first single all-knowing agent worked in demos but hallucinated in production; past a context threshold, extra knowledge degraded performance.
Phase one of Prometheus in New Albany, Ohio is complete and trained Meta's Muse models, including Muse Wonder; Meta is scaling it toward ~1 million GPUs. Training runs hit over 85% goodput for pre-training and over 90% for RL, with mean time between interruptions of ~47k GPU hours, since raised past 54k.
Muse Glimmer runs at practical speeds on a Mac or PC with a single consumer GPU at 4-bit precision, and Meta says it leads comparably-sized open models on tool use. Training placed agents inside simulated digital lives with 10–14 contacts and weeks of messages, calendars and documents.
Meta's standalone Facebook Groups app Forum was built by a small team that ran an AI agent through planning, implementation, build, test and verification, with engineers supervising rather than driving each step. Success was measured in verified, production-ready changes that build, pass tests and survive review — not code volume.
Muse Spark Open is an 816B mixture-of-experts model with 42B active parameters and a 1M token context window, scoring 39 on the Artificial Analysis Intelligence Index — the #1 US open-weights model. Meta trained it on NVIDIA GB300 NVL72 for agentic coding, and it adds audio and video perception to image and document understanding.
Meta's 2026 GEM moves from per-domain optimization (Facebook, Instagram, Reels) to one central foundation model, enabling 2.5x model complexity. It delivered 18.9% conversion and 12% CTR improvement across Meta's ad surfaces.
IcePack is Meta's first liquid-cooled network rack, paired with 102.4T IceCube switch blades (128×800G, Broadcom Tomahawk 6C and NVIDIA Spectrum-6 variants) running FBOSS. Meta also open-sourced its MetaRoCE RDMA transport and launched OSPA with Microsoft and Alibaba.
Meta says the Muse Wonder series delivers state-of-the-art results on personal and professional agentic tasks, memory, and coding, trained at its gigawatt-scale Prometheus data center. Muse Wonder is available via the Meta Model API and Muse Code; Muse Wonder Full is limited to Muse Code subscribers.
Meta built a system that pools locally idle inference capacity across regions and accelerator types into a shared opportunistic pool, improving ads recommendations without ordering more capacity. Two elasticity levers — Rank More Ads (RMA) and Serve Better Models (SBM) — expand optional work while physical and logical isolation plus value-aware load shedding protect baseline latency.