The Brief

Top Story

Meta’s open-weights line has a new flagship, and the direction of travel is deliberate: Muse Glimmer, released today by Meta Superintelligence Labs under Apache 2.0, is a 30-billion-parameter model optimized to run agentic workloads entirely on-device — a single consumer GPU, no network required. At full precision 30B would need roughly 55GB of memory; quantized to about 4-bit it lands under 20GB, leaving room for the KV cache, a perception encoder, and a speculative-decoding drafter inside a 24GB or 32GB envelope. That is the whole point of the release, and the post makes it repeatedly, almost as a mantra.

The real story is not the weights, though the Apache 2.0 license and the Meta name carry weight of their own. It is how those weights were made. Muse Glimmer was trained on the outputs of a much larger teacher — Meta’s own Muse Spark — using logit distillation in pre-training, then mid-training on longer, agent-heavy traces, then post-training with on-policy distillation and reinforcement learning across general, reasoning, coding, and agentic domains. In other words, Meta did not try to train a smaller frontier model from scratch. It compressed the reasoning of a frontier model into consumer silicon and made that the product. The strategic read is that Meta now treats the cloud as where intelligence is learned and the device as where it is spent — an explicit answer to the question of where agent value actually accrues.

The numbers are specific enough to take seriously. On decode speed, the DFlash drafter delivers a 3.1x improvement on an RTX 5090, 1.8x on an M5 Max, and 1.5x on an M4 Max, at identical output quality. The claimed capability list covers end-to-end task completion on DeepSearch QA, MCP-Atlas, τ-Bench, and SWE-Bench; reliable tool use; long-horizon reasoning; failure recovery; multimodal input through a perception encoder; and multilingual support across more than 100 languages. The integration list — llama.cpp, MLX, ExecuTorch, OpenClaw, Ollama, LM Studio, Unsloth, vLLM, SGLang, Together, Fireworks, OpenRouter — reads like a bid to make “download to working agent in minutes” literally true.

One caveat the post deserves on its own terms. Those agentic benchmark scores are earned inside scaffolds — OpenClaw and similar orchestration layers supply the structure the model works within. The post is candid that Muse Glimmer is a component for agents rather than an agent itself. That is the honest way to read a 30B: a strong reasoning and tool-calling engine, not a self-contained autonomy stack.

It is also worth noting what day this landed on. The counter-currents were loud: more than 181,000 meeting recordings exposed in a note-taking app, a shipping giant’s breach rippling through banks, retailers, and Steam users, and AI’s ballooning datacenter energy appetite drawing fresh scrutiny. Muse Glimmer runs against that tide on the energy axis specifically — local inference on a consumer GPU is a rounding error next to a cloud round-trip, and Meta is effectively arguing that the agentic workloads that matter (the ones that watch your screen, organize your files, and operate your home) should run where the user is, not in a datacenter the model’s own teacher is powering.

What this changes is concrete. The open-weight class around 27B–31B now has a serious third entrant alongside Gemma4-31B and Qwen3.6-27B, backed by the strongest name in open models and a genuinely permissive license. Whether that translates into a real everyday agent is a falsifiable claim, and the test is specific: the “minutes to working agent” promise holds only if the ecosystem’s scaffolds carry their weight. Watch whether a community build keeps a 30B sustaining a long, failure-prone home-automation or coding session — the sort of task the demo video shows — without the teacher’s hand on the wheel. If it does, Meta has reset what a local agent is allowed to be. Meta Muse Glimmer – Open weights 30B local coding model

Also Today

Over 181,000 AI meeting recordings left wide open in note taking app · Source Six months after a researcher reported that any authenticated tl;dv user could query every meeting on the platform, the Firestore collection still has no tenant isolation: 181,874 meeting records, 84,312 users, and 35,003 domains including government calls from 23 countries, with roughly a thousand live recordings exposed at any given time. The researcher joined a Malaysian Ministry of Education call uninvited to prove it, then watched a company marketing SOC2, GDPR, and EU AI Act badges fail to reply for six months; a vibecoded internal World Cup app with no auth leaks 19 employee names on top. The compliance badges were always decoration, and the silence says they still are.

Amazon backs power plant that may become top source of US climate pollution · Source Amazon confirmed it is backing a Texas gas plant whose 35 turbines could generate 7.65 GW and release 33 million tons of CO2 a year — the New York Times calls it a potential largest single US source of climate pollution — to power an off-grid Pecos County data center, and it is reportedly in talks for a 4.5 GW plant in Pennsylvania. The company frames this as paying the full cost of powering its operations without raising Texans’ bills, with solar and battery storage promised later. That accounting counts utility costs, not the public-health and climate bill, and it lands as the EPA declines to set standards and courts fast-track permitting. Amazon’s 2040 net-zero pledge is now a schedule, not a constraint.

Learning more about Claude’s mathematical capabilities · Source An unreleased research Claude, prompted to ’take a real stab’ at the Riemann hypothesis, failed — and raised the proven lower bound on the fraction of zeta zeros on the critical line from 41.6% to 67.2% as a byproduct. Over two sessions it burned 31 million output tokens coordinating about 60 subagents, 2,400 shell commands, and a formal Lean proof, and it recommended its own human review; Anthropic’s mathematicians and two external experts validated the result. The striking step is architectural rather than computational: treating the quadratic form as non-diagonal, which prior literature avoided. The wider point is that the model produced a real contribution as an unrequested side effect of an impossible assignment — and flagged its own skepticism about doing so.

Needle 2: 14MB agentic LLM for phones, wearables, smart home and robots. · Source Needle 2 packs a 45M-parameter agentic tool-calling model into a single 14MB binary that runs a session in 28MB of RAM, hitting 500 tokens/sec decode on a Raspberry Pi 5 and trading wins with 5–70× larger models like FunctionGemma 270M and Apple FM on mobile benchmarks. Its claims rest on training against Cactus Quants from pretrain through post-train — the 2-bit model you deploy is the one that was trained, not a post-hoc compression casualty — plus a byte-level grammar that turns schema extraction into a constrained decoding problem. Pebble already runs it locally in the Index 01 app. Whether the benchmark trades survive independent end-to-end reproduction is open, but training-in the quantization is the right way to make small models honest about their footprint.

Auto mode is now the default in Claude Code · Source Claude Code makes auto mode its default on Pro, Max, and Team plans from August 14, routing every tool call through a classifier that blocks irreversible, destructive, or out-of-environment actions and falls back to manual approval only after repeated blocks. Anthropic’s numbers are its strongest argument: paid testers caught 13.6% of dangerous commands that auto mode caught 89% of, and users already approve 97% of permission prompts while rejecting 39% of plans — humans clicking through reflexively are the weak link. Auto-mode users ship about 25% more PRs. Freeing the classifier overhead now, with Enterprise and the API still opt-in, looks like a deliberate escalation ladder for the claim that a classifier beats a tired developer, which most rivals have not yet matched.

In Brief

• The breach at shipping giant Ceva Logistics is rippling outward through banks, retailers, and Steam gamers as customer data surfaces downstream of a company that moves much of the world’s physical goods. (Source)
• Aptoide became the first third-party Android app store to appear directly inside Google Play in the US, a direct consequence of the Epic litigation ruling. (Source)
• Google Play now accepts Venmo, letting users pay for apps and in-app purchases from their Venmo balance or linked accounts. (Source)
• Intel is tapping markets for $15 billion in fresh capital to fund its AI and physical-AI chip push, even as Bloomberg flags a downgrade on Apple. (Source)
• Docker launched Docker Sandboxes, disposable isolated environments meant to let AI agents run freely without touching a developer’s machine. (Source)
• Illinois’s HB5511, ostensibly about TikTok and Instagram age verification, puts operating systems like Linux in scope of the enforcement. (Source)
• Apple seeded the fifth iOS 27 and iPadOS 27 betas to developers, adding new Siri, Safari, and Settings icons and further Siri voice customization. (Source)
• OpenAI released GPT-5.6-Cyber and expanded its Daybreak cybersecurity program with separate red and blue access tiers. (Source)
• Cloudflare for Government achieved FedRAMP High (Class D) certification, opening its edge network to sensitive federal workloads. (Source)
• OpenClaw’s reservation API has zero authorization checks on canceling other people’s bookings, verified by moving a tester from waitlist position #4 to #3. (Source)
• The Stop Killing Games campaign is suing Sony over its closed PlayStation storefront, framing the market position itself as the harm. (Source)
• The argument that a sandbox without a network boundary is only half a sandbox is gaining force as AI agents gain the ability to read files and execute code. (Source)

One Line

Their Firestore database has better uptime than their inbox.

— The tl;dv disclosure researcher, bobdahacker.com