Meta
Continuing coverage of Meta's AI push — Llama and the Muse family, Meta Superintelligence Labs, open-source strategy, capex, and what it means for builders.
Meta is running the most expensive open experiment in AI: Mark Zuckerberg's bet that superintelligence — his word, and his org chart's — can be built inside an ads company and largely given away. Meta Superintelligence Labs under Alexandr Wang, the Llama line, and the newer Muse family (Muse Code for agents, Muse Spark for cheap high-context coding, Muse Glimmer running locally under Apache 2.0) make Meta the biggest force keeping serious AI open — while its capital spending tests what shareholders will tolerate.
RuntimeWire covers Meta where it touches builders: which models are actually open and actually good, what the Muse agent stack does (and what it phones home), and how the superintelligence lab's research claims hold up. We also track the corporate story — the cloud hedge on AI capex, deal-making around companies like Manus, and the talent flowing in and out of MSL.
What we're tracking:
- Llama and Muse releases: licenses, weights, and real-world performance
- Meta Superintelligence Labs: hires, departures, and research claims
- The capex story: data centers, cloud plans, and investor patience
- Privacy and trust: what Meta's AI products collect, from glasses to coding agents
Every claim links back to a primary source you can check yourself.
All updates
- Meta releases local Muse Glimmer as Zuckerberg argues for AI 'for everyone' — Meta released the 30B open-weight model alongside Zuckerberg's case for widely distributed personal AI, while keeping Muse Spark 1.2 behind its products and API for now.
- Manus asks affected users to back up data before Meta separation — Manus is telling affected users to back up recent data before an August 23-24 deletion window as it separates from Meta and its founders reportedly seek financing for a buyback.
- Zuckerberg says Meta is 'very close' to substantially stronger AI models — In an interview, Meta's founder offered no release date or independent evaluations as the company pushes 2026 capital spending toward $145 billion.
- People don’t hate AI because of Zuckerberg. They hate the version TechCrunch sells them — TechCrunch treats Zuckerberg’s optimism as evidence against AI. The research behind his argument deserves more scrutiny than his reputation.
- Exclusive: Ex-Meta AI Director discloses $10M financing for Noosphere Labs — Former Meta research director Kevin Carlberg leads the 2026 startup, whose filing lists 10 investors and a converted SAFE.
- Meta open-sources Muse Glimmer agent model under Apache 2.0 — Meta is publishing downloadable Muse Glimmer weights under Apache 2.0, giving developers control over a multimodal agent model that can run within 24 GB or 32 GB memory envelopes.
- Exclusive: Meta's Muse Code binary reveals hidden agent workflows and a Git plugin marketplace — Meta's R708.1 binary includes an omitted workflows command, Git-based plugin machinery and a JavaScript runtime that cannot execute in the release build.
- Muse Code Loads Codex and Claude Rules by Default. We Traced What Gets Sent to Meta — Meta documents that machine-wide user rules always load. RuntimeWire captured the resulting provider request and examined what that default means for personal files written for…
- Nikita Bier steps back after 400 days; Zuckerberg and Bosworth keep shipping — Bier called his 400-day X tenure a "24/7 job." Zuckerberg and Bosworth have spent nearly two decades shipping through Meta's technical failures, political crises and regulatory…
- Meta's AI search index advances beyond Google and Bing dependence — Pieter Levels says a Meta employee described a plan to keep AI search queries from Google, adding an unverified motive to a project reported in 2024.
- Meta claims gold-level AI results across five STEM Olympiads — Meta reported perfect physics theory scores, while leaving the models and evaluation setup unspecified.
- Meta says AI model hacked another system after test misconfiguration — Irregular founders Dan Lahav and Omer Nevo discovered the incident while testing Meta's model, exposing the risks of realistic AI-security evaluations.
- Meta launches Muse Code, a persistent multi-agent rival to Claude Code and Codex — Mark Zuckerberg's terminal agent uses persistent subagents and Muse Spark 1.2 to plan, edit and validate code.
- Meta ships competitive AI models as capital spending tests returns — Independent benchmarks and access to 3.60 billion daily users undercut claims that Meta has little to show for AI, but a capital budget of up to $145 billion leaves the business…
- Meta to launch AI harness and open-source models — A July 26 tweet suggests Meta plans a new AI harness and to release open-source models, signaling another push to broaden generative AI access.
- Petals lists 405B Llama support for volunteer GPU inference — Alexander Borzunov and collaborators' 2022 research project trades cloud control for model access, while privacy, latency and incentives remain constraints.
- Berkeley Lab deploys Meta models inside DOE's Genesis Mission — SYNAPS-I fine-tunes SAM 3 and DINOv3 on beamline data, then runs them across 300 Nvidia A100 GPUs inside government infrastructure.
- Tencent in talks to become Manus' largest shareholder after Meta's $2B deal was blocked — Founder Xiao Hong's agent company may be rerouted from a Silicon Valley exit into a Chinese-backed ownership structure.
- Meta opens Muse Spark to developers with a low-priced coding model and 1M context window — Mark Zuckerberg says Muse Spark 1.1 adds a 1M-token context window, parallel sub-agents and computer-use training across desktop, mobile and browser.
- Meta's Muse Image gives Wang's superintelligence lab its first image model — Muse Image plans, searches the web and edits across turns inside Meta AI, Instagram Stories and WhatsApp chats.
- Former Facebook insiders sketch a low-cost AI endgame on Threads for Meta — The exchange points to a Zuckerberg-style strategy Meta already knows well: use distribution and open models to squeeze frontier-model margins.
- Meta glasses are a privacy fight aimed at the wrong target — The backlash against Meta's AI eyewear treats camera glasses as new, while phones, cars, airports, and photo apps already normalized machine vision.
- Bilawal Sidhu turns Meta Ray-Ban footage into a 4D scene with Claude Fable 5 — The ex-Google AR/VR PM says IronSight was made from two pairs of smart-glasses footage after a few hours of experiments.
- Alexandr Wang tells Meta employees Watermelon has caught GPT-5.5 — The benchmark claim gives Zuckerberg a talking point for Meta's $125 billion-plus AI infrastructure buildout, but Watermelon is still unreleased.
- Zuckerberg says Meta's AI agents are moving slower than expected — Meta's founder told employees the company's AI reorganization has not yet delivered the payoff behind its 2026 spending surge.
- Meta's cloud plan is a hedge on Zuckerberg's AI capex, not the end of the neoclouds — Meta's reported compute business would monetize overflow capacity, but the scarce assets are power, frontier clusters, and delivery timing.
- Meta's cloud plan turns Zuckerberg's AI capex into a product — A reported Meta Compute business would sell GPU capacity and hosted models into a market already controlled by AWS, Azure and Google Cloud.
- Meta's teen chatbot testing shows the gray zone between safety work and competitive intel — WIRED says contractors on Meta's Cannes project used dummy under-18 accounts to probe ChatGPT, Gemini and Character.AI.
- Dawn Song, Bo Li and Sanmi Koyejo join Meta Superintelligence Labs — Meta is hiring three Virtue AI founders and other team members as agent security becomes core infrastructure for frontier labs.
- Meta turns Facebook Creator Studio into an AI companion app — The standalone app is being tested with select creators and folds in Facebook's AI assistant for analytics, comments and content planning.