Metamate

Metamate is Meta Platforms' internal enterprise AI assistant, available exclusively to company employees (roughly 74,000 people per the 10-K 2024 report). Rolled out in 2024 (Financial Times report), it is a key element of Zuckerberg's 'AI core to how we work' strategy. The product name 'Metamate' comes from 'Metamates' - the internal term for Meta Platforms employees used by Zuckerberg since 2022 (when Facebook Inc. rebranded to Meta Platforms Inc.). This is a deliberate reference: an AI tool for 'Metamates' helping them in their daily work.

Primary use cases (per public sources FT, Fortune, Business Insider, Global Tech Council): (1) Search across internal systems (Metawiki, internal document repositories, JIRA, Workplace, messages in internal chats). (2) Document and conversation summarisation - offloading employees from rarely-read but critical internal resources. (3) Drafting business documents: performance reviews, planning docs, internal memoranda, project proposals. (4) Coding support (supplemented by a separate tool Devmate dedicated to engineers). (5) Translations (Meta operates globally in dozens of languages). (6) Research automation - synthesising results from multiple internal and external sources.

Model architecture: Metamate is multi-model orchestration - the system chooses an AI model per task. The base stack is Meta's own Llama 3.x and Llama 4 models (Llama 4 since April 2025, Mixture-of-Experts architecture). Since December 2024 (Fortune report), Meta announced that Metamate is also supported by OpenAI's GPT-4 via a private API - for tasks where Meta's own models do not yet match frontier competitors. Business Insider (December 2025) reports that Meta employees use a mix of models from OpenAI, Anthropic (Claude), Google (Gemini), and their own Llama - Meta strategically avoids dependency on a single external vendor. This matches an internal 'best model for the task' philosophy. Devmate (dedicated to developers) is a separate variant optimised for code completion and developer support.

Infrastructure and scale: Metamate is hosted on Meta's internal stack - Meta data centres (Menlo Park, Prineville, Los Lunas, Odense DK, Lulea SE, Singapore, Los Angeles LA), inference on a mix of Nvidia GPUs (H100/GB200) + Meta MTIA v1/v2/v3 (Meta Training and Inference Accelerator - Meta's own AI chip designed from scratch). Per Meta's 10-K 2024, AI infrastructure capex in 2024 was ~$40 billion, with a 2025 plan of $60-65 billion (raised to $70-72 billion in H2 2025). A significant portion of this infrastructure supports Llama training, but a key portion of the compute is dedicated to inference (production usage) - including Metamate.

Competition and context: Metamate is Meta's counterpart to Google Moma / Duet (Google Duet for Googlers), Microsoft Copilot for Enterprise (internal), Amazon Q Business (internal at Amazon), Anthropic's internal Claude tools. A key strategic difference: Meta consistently positions itself as an open-source AI player (Llama under the Community licence, publicly available), while simultaneously developing its own closed enterprise assistant. Metamate is a 'dogfooding' product - Meta tests its own Llama models on the most demanding use cases (74k employees, hundreds of thousands of daily queries) before public releases. Significant implication: Meta gathers real-world enterprise AI usage data that public Llama does not have - giving it a competitive edge in the race against Google, OpenAI, Anthropic.