AI Briefing
KOSign in

Apple's 20B 'Core Advanced' Model Uses ~2GB RAM on Macs

·2026.10.25 17:17

Key point

Apple's 20B 'Core Advanced' model runs on macOS 27 using approximately 2.0–3.4 GiB of RAM by storing weights on flash, consuming about one-third of the energy of comparable open models.

1 / 10

Details

  • Apple's 20B parameter 'Core Advanced' model runs on macOS 27 using approximately 2.0 to 3.4 GiB of RAM (including safety models), compared to 15 GiB for the Gemma 4 26B A4B model.
  • The model consumes about one-third of the energy of the best open builds on math tasks (465 J vs 1,273 J for Qwen3 4B) while keeping the GPU idle and running on the Neural Engine at ~4W.
  • Architecture relies on sparse decoding with 1-4B active parameters per prompt, using a Gemma 3-based tokenizer and weights stored in a 4.95 GB flash file loaded in 294,912-byte blocks.
  • Performance is limited for complex reasoning: it solved 78/100 GSM8K math problems (vs 89-94 for open models) and 67.0% on MMLU, making it suitable for short tasks like summarization and classification rather than coding or math.
  • The model is built in collaboration with Google, using Gemini outputs for refinement and cloud TPUs for pre-training, but lacks developer-facing adapters and model pinning APIs in macOS 27.
applemoeon-device-aigoogle-geminineural-enginemacos-27core-advanced
Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.

Related

Apple Newsroom (Korea)Apple begins Korean language support for Siri AI with iOS 27.2 updateRedditOpenBMB Releases MiniCPM-V 4.7: 35B Multimodal Model Runs on Mobile with ~3GB MemoryHuggingFace BlogLiquid AI Releases LFM2.5-24B-A2B: A 24B-Parameter Reasoning Model with 2B Active Parameters for On-Device AgentsGeekNewsMacBook Air에서 1시간 분량 오디오를 약 13초 만에 전사RedditOpenBMB Releases MiniCPM-V 4.7 35B-A3B, First Open-Source MoE Vision Model with Smooth On-Device InferenceTLDR AITony Fadell: First Wave of AI Gadgets Failed by Ignoring Real User Needs