The LLM included in macOS 27 isn't widely covered in articles, so I decided to take a closer look: https://t.co/7m33xszKVb Can the AI Built into Mac Be Used for Work? An Engineer’s Review of Apple’s “fm” [macOS 27 / fm Command / Local LLM] Usagi Hacker’s AI Lab by CodeRabbit #AISummary The Capabilities and Use Cases of macOS 27’s Local LLM: “fm Command” 🔳 Apple Has Integrated a Local LLM by Default The video introduces “fm command,” a feature in macOS 27 that enables access to Apple’s local LLM. It has drawn attention as a mechanism allowing AI execution on Apple Silicon Macs without relying on cloud services. 🔳 Offline, Free, and Unlimited Usage The fm command operates without an internet connection and incurs no additional fees regardless of usage frequency. Since processing occurs locally without transmitting data externally, it offers significant privacy advantages. 🔳 Requirements and Initial Setup The video explains that Apple M3 chips or later, along with at least 12GB of RAM, are required. First, accept the terms of service using “fm license,” then check model availability with “fm available.” Once the model is activated, usage can begin. ListItemIcon Main Features of fm Command “fm chat” enables conversational AI, while “fm respond” handles single queries. It also supports OCR for text extraction from images, barcode reading, structured output in JSON format, and API server functionality. ListItemIcon External Integration via OpenAI-Compatible API “fm serve” launches a local API server. Through an OpenAI-compatible interface, external tools like OpenCode can invoke fm and integrate it into AI agents or automation systems. ListItemIcon Integration with Python and Shell Scripts Combined with Python SDK or shell scripts, fm enables automated tasks such as translation, OCR, and file organization. Specifying output formats via JSON schemas makes it easier for programs to process results. ListItemIcon Major Limitation: 8192-Token Context Window The context window is limited to 8192 tokens, making it unsuitable for long-form processing or large-scale coding tasks. The video compares it to GPT-3.5 Turbo from around 2023, noting significant constraints compared to current high-performance LLMs. ListItemIcon Outdated Knowledge and Hallucinations Testing in the video reveals the model’s knowledge is frozen around October 2024. It cannot autonomously search for up-to-date information and may produce factual errors or hallucinations. However, appropriate responses were observed in some cases—such as correctly rejecting fictional narratives. ListItemIcon Coding Validation with OpenCode Configured as a sub-agent in OpenCode, fm successfully implemented basic add/delete functionality for an HTML To-Do app in about 20 minutes. However, the UI remained minimalistic, and issues persisted with Japanese input. ListItemIcon Potential as a Sub-Agent Given its strict context limitations, fm is more practically suited as a sub-agent handling limited tasks—such as translation, text extraction, or simple code generation—rather than as a primary AI. ListItemIcon Potential for Bulk Data Automation Ideal for processing hundreds to thousands of images via OCR or translating long documents by splitting them into segments. Since there are no usage-based fees, it offers clear advantages when combined with programs for repetitive tasks. ListItemIcon Overall Assessment at This Stage While it does not yet match the performance of the latest cloud-based LLMs, its free and offline nature holds value. For developers seeking to experiment with small-scale automation or AI agents, it presents an intriguing option.