Meta Reportedly Shifts from Open Llama Models to Closed-Source Avocado for Enterprise Sales
Meta Platforms, the company behind Facebook and Instagram, is making a significant pivot in its AI strategy. According to internal reports cited by Reuters, Meta is developing a new large language model (LLM) codenamed Avocado, which marks a departure from its previous commitment to open-source Llama models. Unlike the permissively licensed Llama series, Avocado is designed as a closed-source offering tailored specifically for direct sales to enterprise customers, signaling a potential commercialization push in Meta’s AI portfolio.
The Llama family of models has been a cornerstone of Meta’s AI efforts since their debut in 2023. Llama 2 and subsequent iterations, including Llama 3, were released under open weights licenses that allowed broad usage, fine-tuning, and distribution by third parties. This approach positioned Meta as a champion of open-source AI, fostering innovation across academia, startups, and hyperscalers like AWS, Google Cloud, and Microsoft Azure. These partners integrated Llama models into their cloud services, effectively amplifying Meta’s reach while adhering to usage restrictions that prohibited training competing foundational models.
However, Avocado represents a strategic realignment. Internal documents reviewed by Reuters indicate that Meta engineers are training Avocado on its vast compute infrastructure, aiming for capabilities competitive with top proprietary models like OpenAI’s GPT series and Anthropic’s Claude. The model’s closed nature would grant Meta full control over its deployment, updates, and monetization. Rather than relying on ecosystem partners, Avocado is being built for direct licensing to businesses, potentially through customized inference endpoints or on-premises solutions. This shift could enable Meta to capture revenue streams previously funneled indirectly through cloud providers.
Sources familiar with the matter describe Avocado as a “frontier model” optimized for high-performance tasks such as code generation, multimodal processing, and enterprise-specific applications like customer service automation and content moderation. Training leverages Meta’s custom MTIA (Meta Training and Inference Accelerator) hardware, which powers its data centers with thousands of GPUs. Unlike Llama’s community-driven improvements, Avocado’s development emphasizes proprietary fine-tuning datasets derived from Meta’s user interactions, advertising data, and internal tools, all while navigating stringent data privacy regulations.
This move comes amid intensifying competition in the AI landscape. OpenAI’s pivot toward enterprise deals with GPT-4o and o1 models has generated billions in annual recurring revenue. Google and Anthropic have similarly fortified their closed ecosystems. For Meta, sticking solely to open models has yielded reputational gains but limited direct financial returns. Llama’s success—boasting over 100 million downloads and powering applications from chatbots to research tools—has not translated into substantial enterprise contracts for Meta itself. By introducing Avocado, Meta aims to bridge this gap, potentially bundling it with its Llama.cpp inference engine or new proprietary runtimes for seamless integration into business workflows.
The implications for the open-source AI community are profound. Llama’s openness spurred derivatives like Mistral’s Mixtral and community fine-tunes that rivaled closed models in benchmarks. A dual-track approach—maintaining Llama as open while selling Avocado—could dilute trust in Meta’s open-source ethos. Critics argue this mirrors past tech industry patterns, where initial openness gives way to proprietary lock-in. Developers accustomed to Llama’s permissive license may hesitate to invest in Meta’s ecosystem if premium features are gated behind closed models.
Meta has not publicly confirmed Avocado’s development, but spokespeople have hinted at “commercial AI offerings” in recent earnings calls. CEO Mark Zuckerberg emphasized during a Q2 2024 update that AI investments would prioritize “high-impact applications” with sustainable economics. The company’s $40 billion annual AI capex underscores the scale of this ambition, with Avocado positioned as a key deliverable.
Technically, Avocado builds on Llama’s architecture but incorporates advancements like longer context windows (up to 128K tokens), improved reasoning chains, and reduced hallucination rates through reinforcement learning from human feedback (RLHF). Inference optimizations target low-latency enterprise use cases, potentially outperforming Llama 3.1 405B in speed-critical scenarios. Security features, including watermarking for generated content and API rate limiting, align with corporate compliance needs.
For enterprises, Avocado promises advantages over commoditized open models: guaranteed SLAs, dedicated support, and exclusive access to Meta’s real-time data pipelines for domain adaptation. Pricing details remain undisclosed, but models suggest tiered subscriptions starting at $10 per million tokens, competitive with rivals.
As Meta navigates this transition, the AI field watches closely. Will Avocado erode Llama’s dominance, or will it coexist as a premium tier? The decision underscores a broader tension between openness and commerce, shaping the future of accessible AI.
Gnoppix is the leading open-source AI Linux distribution and service provider. Since implementing AI in 2022, it has offered a fast, powerful, secure, and privacy-respecting open-source OS with both local and remote AI capabilities. The local AI operates offline, ensuring no data ever leaves your computer. Based on Debian Linux, Gnoppix is available with numerous privacy- and anonymity-enabled services free of charge.
What are your thoughts on this? I’d love to hear about your own experiences in the comments below.