Meta just dropped a new open-weight AI model called Muse Glimmer, 30 billion parameters, Apache 2.0, runs on a single consumer GPU. Zuckerberg published a manifesto to go with it, arguing open weights are how the US catches up to Chinese labs. Bitcoin is drifting near $63,700 after July CPI came in at 3.4%, right on the nose. Public miners have quietly added $1.78 billion of selling pressure. A volunteer Bitcoin Red Team used AI to find nearly 5,000 potential vulnerabilities across 390 open-source Bitcoin projects in about 30 hours. And the OCC is now taking public comments on how it'll actually license stablecoin issuers under the GENIUS Act. Let's get into it.
So Meta shipped Muse Glimmer yesterday, and this is a real strategic pivot. It's a 29.6 billion parameter dense model, 128K context window, released under Apache 2.0, with weights on Hugging Face. Text and image input, tool calling, and it's built explicitly to run locally. Full precision it needs about 55 gigs of VRAM, but the 4-bit quantized version fits under 20 gigs, which means it runs on a MacBook M4 or an RTX 5090 with room left over for the KV cache and a speculative decoding drafter. Meta claims it beats Google's Gemma 4 31B on a bunch of benchmarks, and third-party evals put it competitive with Claude Haiku 4.5 and close to Gemini 3.5 Flash-Lite. Zuckerberg published a long essay to frame it. His argument: the US is losing the open-weight race to Chinese labs like Alibaba's Qwen and Moonshot's Kimi, and Meta wants to be the American answer. He's pushing Washington to reduce friction on open-source AI and pitching a governance structure with independent directors approving safety criteria. Read the strategy honestly. Meta is behind OpenAI and Anthropic on enterprise revenue. Yann LeCun got replaced by Alexandr Wang. So instead of fighting a losing battle on frontier closed models, they're going bottoms-up. Personal AI, on-device agents, developer ecosystem, function calling with schemas, multi-step reasoning, all running without a cloud bill. The bet is that if superintelligence gets distributed instead of concentrated, Meta wins by being the platform underneath it. Whether that's principled or just pragmatic depends on how cynical you are, but the model itself is genuinely impressive for what runs on a laptop.
While Meta reboots, AI agents are quietly landing in actual production systems, and the case studies this month are worth paying attention to. The US Army Human Resources Command just deployed Salesforce Agentforce in an IL5-authorized environment. That's the first Department of War entity running autonomous AI agents on controlled unclassified information. It's a 24/7 system supporting soldiers, veterans, and families, handling routine inquiries, summarizing case histories, and pulling policy from approved Army sources. Complex decisions still route to humans. Projected savings: about $6 million a year, over 55 million agent conversations per month at full scale, 1,500 automated case summaries per day. LendingTree meanwhile built a multi-agent mortgage assistant on Amazon Bedrock. Three agents, orchestrated by LangGraph, running on ECS and Fargate. A supervisor handles intent, an education agent uses RAG grounded in real documents, and a matching agent calls LendingTree's actual API suite for offers and rates. Average sessions run 10-plus messages over about 9 minutes. This is real multi-turn work in a regulated industry with PII protection and content filtering baked in. CarMax expanded its Sierra voice agent deployment for inbound sales calls. Wyndham Hotels hit a 62% automation rate with Five9, automating 40,000 password resets monthly and dropping call abandonment under 1%. And KT finished building NH Nonghyup Bank's AI contact center in Korea, expanding AI-handled tasks from 45 to 180 with 97% speech recognition accuracy. The pattern here matters. These aren't demos. They're not chatbots bolted onto FAQ pages. They're production agent systems with orchestration, tool calling, guardrails, and measurable ROI. The agentic era isn't coming. It's shipping.
Some real infrastructure movement on Bitcoin this week. Lightning Labs released the alpha of Wavelength, a non-custodial API toolkit for embedding Bitcoin payments into apps without running your own node or channels. It supports on-chain Bitcoin, Lightning via atomic swaps, and an Ark-style off-chain settlement layer for fast cheap transfers. The interesting part: it's accessible to AI agents through the Model Context Protocol, meaning agents can hold Bitcoin balances and pay for services in fractions of a cent. Fees are 1 basis point on Lightning during alpha, plus routing. SDK is open source, testnet now, mainnet by invitation. Stablecoins over Taproot Assets planned later. Blockstream also launched Blockstream Swaps, a trustless cross-layer swap tool for moving between mainnet, Lightning, and Liquid without custody. This comes right after Boltz suspended its non-custodial swap service citing AI-assisted attack concerns. Which is itself the whole vibe of Bitcoin infrastructure right now. Speaking of which, the Bitcoin Red Team, a 16-person volunteer group including Calle and AnchorWatch CEO Rob Hamilton, ran an AI-assisted audit across Bitcoin open-source repos. In about 30 hours they surfaced 4,962 potential issues across 390 projects, with roughly 720 flagged as high or critical severity. About 21% reproducible so far. That's roughly one critical exploit per person per hour. This is after the Coldcard hack that took over $100 million. The takeaway is uncomfortable but honest: AI is now the best code auditor available, and it's finding things everywhere. Bitcoin's open-source stack is going to get hardened fast, or attackers with the same tools are going to make everyone pay. The Bitcoin Policy Institute is separately asking frontier AI labs to give Bitcoin developers early access to models for exactly this reason. Sensible ask.
The GENIUS Act is turning into actual paperwork. The OCC opened public comment on the licensing and registration forms for payment stablecoin issuers. Domestic applicants need a full business plan, reserve and redemption policies, biographical and financial reports for every director and principal shareholder. Foreign issuers need Treasury to certify their home regime is comparable to GENIUS standards, which is a real gate. OCC estimates about 125 hours of burden per application. That's bank-charter-level filing work. Comments close September 25. The ABA weighed in saying stablecoin issuers should be treated like financial institutions for BSA and sanctions purposes, but wants IDI subsidiaries to piggyback on parent bank AML programs. Paradigm and the Hyperliquid Policy Center pushed back the other direction, arguing the rules could be read to sweep in validators and protocol developers, which they say Congress never intended. Both sides have a point. Overseas, Standard Chartered-led Anchorpoint launched a Hong Kong dollar stablecoin with HashKey and OSL as authorized distributors. The Bank of England is going to test stablecoin cross-border trade finance in its Digital Pound Lab. Slovenia became the first jurisdiction to register a stablecoin issuer under MiCA. And separately, Nasdaq is acquiring LeveL Markets to push into always-on trading. On the market, Bitcoin drifted to $63,700 after CPI landed at 3.4%. ETF inflows are being offset by public miner selling at $1.78 billion in pressure, plus transaction fees have collapsed to 0.7% of miner revenue, a 10-year low, which is why so many miners are pivoting to AI hosting. Bitdeer just filed for a $1 billion equity raise for AI buildout, potentially diluting shareholders up to 30%. Strategy's Michael Saylor said they'll resume Bitcoin accumulation this year and have a $4.6 billion cash buffer, roughly three years of runway before any forced BTC sales. And it's now been five years since El Salvador made Bitcoin legal tender. The verdict from locals is mixed at best, but for Bitcoin's global profile the experiment moved the needle.
One prediction to end on. Every serious AI story this week, from Muse Glimmer running on a laptop, to Army agents in production, to Wavelength giving agents their own Bitcoin wallets, points to the same endpoint. The frontier is no longer the model. It's what the model can hold, spend, and transact on its own. Bitcoin's Lightning stack is quietly becoming the settlement layer for that world, and most people haven't noticed yet.