META is pivoting back to open source around Muse Glimmer, a 30B open-weight model that runs locally on a MacBook with ~20 GB of RAM and scales to ~55 GB, because cheap private AI is framed as the path to adoption, data capture, and pressure on closed-model margins. The bet is that release cadence plus massive compute — including a planned $210 billion of spend through 1H26 and a $145 billion 2026 capex guide — can turn open weights into a distribution weapon across AI, ads, glasses, and edge inference. The counterpoint is capital intensity and execution risk: the strategy burns cash, depends on Meta shipping repeatedly, and only works if local models plus infrastructure can beat frontier APIs on cost and enough capability for most work. Across the market, falling model prices and rising compute commitments from OpenAI, Anthropic, NVDA, and META are pushing the video’s core thesis: intelligence is getting commoditized, so the winners are the firms with scale, distribution, and hardware control.
META is pivoting back to open source for the third time, reversing its closed-source turn with a 14-page paper in the last earnings report.
Muse Glimmer is the new open-weight open source model: 30B parameters, ~20 GB to run on a MacBook, and ~55 GB for the fully capable largest version.
Muse Glimmer is quantized, has a couple hundred thousand tokens of context, and is positioned as a cost/efficiency-per-token model, not a frontier model.
Local AI is framed as the target state because it is cheap, can use private data, and is efficient for serving answers; the example given is sensitive health data that should not go to a public cloud model.
META says open source is a way to let everyone build on the model and strip margins away from frontier labs; the video calls this “scorched earth 2.0.”
META also open-weighted Muse Spark 1.2, described as a frontier model released about a week and a half earlier and a strong coding model with agents that replicate and split code work in parallel.
The release strategy is aimed at models that do 90% to 95% of the work cheaply, not at maximum intelligence, with Zuck repeatedly arguing local private models will outcompete frontier models.
Using Anthropic or OpenAI APIs at a couple thousand dollars a month, or even a couple hundred, is presented as making Muse Glimmer the better bang-for-buck workhorse versus cloud subscriptions.
Enterprise AI spend is described as ranging from $10 million to hundreds of millions of dollars per year, so lower-cost local/open models are presented as the likely purchase choice for most workloads.
Zuck’s three-pillars framework is: give everyone access to their own private AI model; let invention determine who gets access; and preserve a core balance of power.
The bear case is that META floods the market with cheap intelligence, commoditizes paid subscriptions at frontier labs, cuts their margins and revenue, and buys time for META to catch up.
META is spending aggressively on that strategy: about $210 billion of capex through the first half of next year, a raised 2026 capex guide of $145 billion, and free cash flow this year down 91% YoY.
META is one of the largest holders of NVDA GPUs and NVDA Rubens, second only to Elon Musk, and the biggest proof point for the strategy is release cadence.
In the last 90 days META has released two models and is expected to release two more over the next one to two months; Elon Musk is also trying to ship iteratively the same way.
META’s consumer moat is the other half of the bet: it owns 70% to 80% of the AI glasses market with Luxottica, tripled AI glasses sales in 2025, and says AI chats will personalize ads with real-time preference profiles like “Josh.md” and “Ejas.md.”
The broader market setup is that eight models were released over the last 90 days, including three from OpenAI, Terra cut prices 80% a month ago, and Anthropic said Claude Sonnet will stay at its cheapest price ever indefinitely.