Show HN: Apfel – The free AI already on your Mac

Original Article ↗ Hacker News Discussion ↗

Apple’s new on-device language model in macOS “Tahoe” is being tapped by an open-source CLI tool, Apfel, which exposes it for shell use and as an OpenAI‑compatible local server. Commenters highlight its strengths—instant startup, strong privacy, multi-language support, and usefulness for small, local tasks—while noting real limitations such as a 4K context window, conservative safety guardrails, weak math and coding performance, and quirks in non‑English usage. There is also concern over securely exposing localhost AI APIs and frustration with Apple Intelligence’s OS and UX trade-offs, alongside broader optimism that local LLMs will become an important alternative to cloud-based AI.

Requirements & installation

  • Works only on Apple Silicon Macs running macOS 26 “Tahoe” with Apple Intelligence enabled; fails on Sequoia and earlier due to missing FoundationModels.framework.
  • Enabling Apple Intelligence triggers a separate, large model download; Apfel itself is a small binary (~4–15 MB).
  • Some users are reluctant to upgrade to Tahoe or to enable Apple Intelligence at all, despite interest in the tool.
  • A Homebrew tap PR was submitted so it won’t install on unsupported macOS versions.

Capabilities & limitations

  • Wraps Apple’s on-device Foundation Model with a CLI and optional OpenAI-compatible server.
  • Hard 4K token combined context window; repeatedly cited as the main limitation, especially for coding, logs, and “sub‑agent” use with larger models.
  • Said to be “not made for conversation”; better suited to short prompts, shell scripts, quick facts, simple data analysis, JSON → sentence, etc.
  • Supports multiple languages (en, de, fr, zh, etc.), but users report quirks (e.g., trouble switching German Du/Sie, defaulting to German decimal notation).
  • Built-in safety/guardrails are very strong; sometimes refuses surprisingly benign tasks or feels like “Siri” in cautiousness.

Model quality & behavior

  • Users report high non‑determinism and frequent mathematical/timezone mistakes, plus occasionally messy formatting.
  • Some find it hallucination‑prone on “what do you know about X?” questions and on date/time arithmetic.
  • Others report surprisingly good performance on specific structured tasks (e.g., local pricing/cost prediction backtesting), beating both frontier and other local models for their use case.

Privacy, security & local use

  • Runs fully on-device; the FoundationModels API used here has no access to personal Apple account data or Apple’s internal semantic index/RAG.
  • Strong interest in local models for privacy and offline/agentic workflows; debate over whether local is strictly necessary vs. “zero-retention” cloud models and TEEs.
  • Security concern: exposing an HTTP API on localhost can be driven by arbitrary web JS. Apfel’s server is off by default, has optional bearer auth, and was hardened after feedback; new security docs were added.

Ecosystem, UX & positioning

  • Compared against Qwen and other local models: Apple’s model is viewed as small but efficient; concern that it may lag behind rapidly improving 4B–level OSS models.
  • Related tools mentioned: GUI frontends, local STT/TTS, and OS launchers (Alfred, krunner, PowerToys) as good integration points.
  • Some praise the project, simplicity, and local-first approach; others criticize the marketing-heavy landing page and argue it’s “just” a wrapper around an existing Apple API.