Meta's Llama 3.1 release clashes with internal unrest

Meta shipped Llama 3.1, a 405‑billion‑parameter open‑weight model that sits next to Anthropic’s Claude 3 and OpenAI’s GPT‑4o in raw size. The launch pushes Meta’s open‑source AI narrative forward and adds quantized FP8 versions that run on a single 8‑GPU node.

The suite arrives in three sizes—8 B, 70 B, and the flagship 405 B. All three use a dense feed‑forward transformer trained on 15.6 trillion curated tokens. Meta’s license permits synthetic‑data generation but retains branding clauses. Pricing for the hosted API mirrors GPT‑4o, ranging from $3‑9 per million input tokens and $3‑15 per million output tokens, while on‑premise deployment can undercut those rates dramatically.

Employee morale erodes under AI push

An internal livestream of an Applied AI meeting was hijacked by an angry engineer who shouted that he was “the company’s bitch” and demanded that a senior AI executive be called a “piece of shit.” The outburst was captured by WIRED and illustrates a broader malaise inside Meta’s Applied AI unit, which was assembled in March to support the new Meta Superintelligence Labs.

Three current employees, speaking on condition of anonymity, described the unit’s 6,500 engineers and product managers as stuck doing “menial” tasks such as generating puzzle sets to test model reliability. One called the work “literally the gulag.” The sentiment echoes a company‑wide petition signed by more than 1,600 staff demanding a pause to Meta’s internal click‑ and keystroke‑monitoring program that harvests data for AI training. The program now allows a 30‑minute pause and limited exemptions, but the backlash shows how the recent 10 % layoff—8,000 jobs—has amplified stress across divisions.

Meta chief product officer Chris Cox later addressed Instagram employees, describing the environment as “brutal” and likening it to “running a marathon in the middle of a hailstorm.” Cox admitted leaders need to “get in touch with the company again” and warned that AI is “neither god nor devil.” His remarks underscore the disconnect between the technical ambition of Llama 3.1 and the day‑to‑day experience of the engineers building it.

Capital spending fuels ad‑driven AI revenue

In FY 25 Meta reported $198.7 B in Family of Apps revenue, up 22 % year‑over‑year. AI‑enhanced ad tools handled more than $60 B of annualized ad spend through the Advantage+ suite, while video‑generation tools hit a $10 B combined revenue run rate. The company spent $72.2 B on capex and now guides FY 26 capex to $115‑135 B.

Meta’s AI pricing model is the opposite of most AI firms: there are no API fees, subscriptions, or per‑token charges for end users. Every AI capability— from the Meta AI assistant to generative ad‑creative tools— is offered for free. The model monetizes indirectly; free AI features boost engagement, create ad inventory, and drive revenue.

Three model architectures disclosed on earnings calls have delivered measurable conversion gains. A new runtime model launched on Instagram Feed, Stories, and Reels in Q4 lifted conversion rates by 3 %. An incremental attribution feature raised incremental conversions by 24 %, reaching a multi‑billion‑dollar annual run rate within seven months. Those improvements compound because better predicted action rates feed directly into Meta’s ad auction, raising the effective price per impression without adding more ads.

Avocado and the unanswered trade‑offs

Meta’s next generative‑AI effort is codenamed “Avocado.” Insiders tell CNBC the model was expected by year‑end but is now slated for a Q1 2026 release. The project is still undergoing performance testing, and a Meta spokesperson said training is “going according to plan.”

The Avocado push follows a $14.3 B hiring spree in June that brought Scale AI founder Alexandr Wang and his team into Meta. The same month Meta bought a large stake in Scale AI and lifted its FY 25 capex guidance to $70‑72 B. Analysts at KeyBanc note that Meta entered 2025 as an AI winner but now faces “more questions around investment levels and ROI.”

Avocado raises a strategic dilemma. Meta’s historic advantage lay in open‑source Llama models that let the broader community experiment, distill, and fine‑tune. Avocado’s roadmap is opaque, and the company has not disclosed whether it will follow the same open‑weight licensing or adopt a more closed stance. If the model remains open, it could accelerate research on distillation and synthetic data, as the 405 B Llama 3.1 already enables. If it becomes proprietary, Meta risks abandoning the community that helped it close the gap with closed‑lab giants.

The tension between open‑source ambition and the need for a clear ROI on multi‑billion‑dollar capex is the core friction that many external observers miss. Meta’s ad‑driven revenue model rewards incremental prediction improvements, but the cost of training and maintaining frontier‑scale models can outpace the marginal gains unless the models are widely adopted beyond internal ad tools.

What to watch

Track the Q1 2026 Avocado launch and any licensing announcements. Monitor Meta’s capex spend against FY 26 guidance; a deviation could signal a shift in AI priorities. Keep an eye on employee sentiment—especially any further petitions or public outbursts—as morale may affect the speed of model development. Finally, watch adoption metrics for Llama 3.1 in the open‑source community; strong external uptake could pressure Meta to keep future models open, while tepid uptake may push the company toward a more closed, revenue‑focused approach.