Meta opens Muse Spark, Microsoft’s efficient giant, and agents go shopping

Compact Conversations for 2026-10-15: 5 AI stories, ai news worth knowing in just 5 minutes.

[Audio embed placeholder]

The Lead: Meta releases Muse Spark Open

Meta unveiled Muse Spark Open, an 816‑billion‑parameter open‑weight model with 42 billion active parameters and a 1‑million‑token context window, optimized to run on two Apple M5 Ultra Mac Studios via the MLX framework.

Why it matters: It brings leading agentic capabilities to desktop hardware, showing that frontier models can run locally on high‑end Macs and offering developers a powerful, openly licensed tool for coding and multi‑step reasoning.

Source: Techmeme

The Feed

Agentic commerce hits 138M transactions on Base

Chainalysis reported that AI‑agent transactions on Coinbase’s Base blockchain grew from ~1,500 in Q2 2025 to over 138 million in Q3 2026, with more than 87 % of payments under ten cents.

Why it matters: The explosion of micropayments shows how crypto‑stablecoin rails enable AI agents to autonomously purchase compute, data, and tools, hinting at a new AI‑to‑AI economy.

Source: Fortune

Anthropic reports $16.8B sales, positive cash flow

Anthropic posted $16.8 billion in quarterly sales — up nearly 50 % — and achieved positive cash flow, marking its strongest profitability signal yet.

Why it matters: The numbers underscore Anthropic’s path to profitability and bolster confidence ahead of a potential IPO, highlighting the maturing business side of frontier AI labs.

Source: Axios AI+

Microsoft unveils 2.8T‑parameter MAI‑Base‑2 model

Microsoft introduced MAI‑Base‑2, a mixture‑of‑experts model with 2.8 trillion total parameters and 80 billion active, claiming 4×‑12× better pretraining FLOPs efficiency than DeepSeek‑V4‑Pro‑Base.

Why it matters: The efficiency gains suggest a cheaper route to frontier‑scale models, potentially lowering costs for enterprises that rely on large‑scale AI training.

Source: Techmeme

FTC subpoenas OpenAI and Anthropic over consumer harms

The U.S. Federal Trade Commission issued subpoenas to OpenAI and Anthropic as part of an investigation into alleged consumer harms from their AI products.

Why it matters: This marks the first major U.S. regulatory action against the leading AI labs, signaling increased scrutiny over safety, transparency, and market practices.

Source: Axios AI+

One Thing to Try

Set up a lightweight agent that runs a scheduled job — like checking competitor prices — and delivers the result to your chat even with your laptop closed. The Reddit post shows how to use simple scripts and Zapier‑style rules without paying for OpenAI’s $100/month offering.

Sources

Transcript

Host A: Welcome to Compact Conversations, the show that compresses the day’s AI news into 5 minutes.

Host A: Meta released Muse Spark Open today, calling it a highly capable open-weight model from a US company so far. The specs are big: 816 billion total parameters, though only 42 billion are active at any moment, plus a context window of one million tokens. It’s built for coding and agentic work, and the company says it does well on long-horizon tasks and multi-step software engineering.

Host B: [curious] There’s a fun hardware angle here too. Meta says it’s the first model in its class to run on your desktop, using two Apple M5 Ultra Mac Studios with the MLX framework. Apple’s unified memory lets each Mac Studio address up to 512 gigabytes, so two machines together hold the whole model — Meta describes it as two Mac Studios acting like one machine with a terabyte of unified memory. [with a small lift] They demoed it shipping real iOS changes to an open-source Mastodon client and driving DaVinci Resolve end to end.

Host A: [conversational] Weights are on Hugging Face and the usual cloud model catalogs, with hosted inference through providers like Together AI, Fireworks, and Groq. It ships under the Muse Model License with limited commercial terms, and Meta says it added extra safety post-training because open weights carry their own considerations.

Host B: One number to know today: 16.8 billion dollars. That’s Anthropic’s quarterly sales, according to Axios — and the company also reported positive cash flow. [thoughtful] It’s the strongest profitability signal yet from a leading AI lab, and Axios notes Anthropic is eyeing an IPO.

Host A: In the feed, Fortune reports on a Chainalysis study of agentic commerce — AI agents paying for services on their own. On Coinbase’s Base blockchain, transactions through the x402 payment protocol went from about 1,500 in mid-2025 to more than 138 million by the third quarter of this year.

Host B: [curious] Most of it is tiny money — over 87 percent of payments were under ten cents, where card fees would cost more than the purchase. Developers fund agent wallets with stablecoins, and the median time from first deposit to first payment dropped from 73 days to 18 minutes.

Host A: Next, Microsoft unveiled MAI-Base-2, a mixture-of-experts model with 2.8 trillion total parameters and 80 billion active. Microsoft says it beat every open-weight model they tested on held-out evaluations, and the training recipe is four to twelve times more compute-efficient than DeepSeek’s strongest open base model. The model is still in post-training.

Host B: [thoughtful] And Axios scooped that the FTC subpoenaed OpenAI and Anthropic over consumer harms.

Host A: One Thing to Try: set up an always-on personal agent that runs a job on a schedule and delivers the result to your chat, even with your laptop closed. This comes from a builder writing in r/AI_Agents. [conversational] You hand it a plain sentence — once, or every morning — and it does the reading and judgment work for you. The author’s own example: agents watching two competitor pricing pages that only ping when something actually changed. Their team also runs a morning brief with three things that matter, one line each, with links.

Host B: [quick aside] The distinction they draw is useful: automation tools like Zapier move data between apps by rules, but this is the part that needs judgment — read, compare, decide, write. [lighter] Pick one thing you check manually every morning and hand it to an agent instead. The catch: OpenAI’s version of this starts at a hundred dollars a month and isn’t available in Europe, so look for a lighter-weight option to experiment with.

Host A: That’s Compact Conversations for Thursday. More AI news tomorrow. Until then, happy prompting.