Hardware sizing

Can my Mac or PC run local AI?

Pick your machine's RAM, GPU VRAM, and desktop AI workloads to see exact disk downloads, resident memory, KV-cache overhead, and which open-weight speech, summary, search, and autocomplete models fit without starving your other apps.

Estimating...

Calculating fit...

Disk download —
Warm resident RAM —
Combined peak —
50% RAM cap —

Recommended local model stack (swap any row to test)

Role Model & quantization Disk Active RAM Measured benchmark

Quick reference

Recommended local AI stacks by machine tier

Every model size below comes directly from LokalBot's shipped macOS (ModelCatalog.swift, LlamaServer.swift) and Windows/Linux (catalog.rs, sherpa-onnx) catalogs. Screen OCR uses Apple Vision on macOS (0 GB download) and Windows OCR or Tesseract + ocrs on PCs.

Machine tier 50% RAM cap Transcription Summaries & memory Autocomplete Search embedder Total disk
Mac · 8 GB unified 4 GB Qwen3-ASR 0.6B (1.02 GB) or Parakeet v3 (0.60 GB) Qwen3.5 0.8B Q4_K_M (0.53 GB) LFM2.5 1.2B Q4_K_M (0.73 GB) EmbeddingGemma 2 Q8_0 (0.31 GB) ~2.8 GB
Mac · 16 GB unified (Default) 8 GB Qwen3-ASR 1.7B (2.52 GB) + Nemotron 3 (0.20 GB) Qwen3.5 4B Q4_K_M (2.74 GB) Gemma 4 E2B i1-Q6_K (3.85 GB) EmbeddingGemma 2 Q8_0 (0.31 GB) ~9.6 GB
Mac · 24 GB unified 12 GB Qwen3-ASR 1.7B (2.52 GB) + Nemotron 3 (0.20 GB) Qwen3.5 4B (2.74 GB) or Gemma 4 12B (7.12 GB) Gemma 4 E2B i1-Q6_K (3.85 GB) EmbeddingGemma 2 Q8_0 (0.31 GB) ~9.6–14.0 GB
Mac · 32–48 GB+ unified 16–24 GB+ Qwen3-ASR 1.7B (2.52 GB) + Nemotron 3 (0.20 GB) Qwen3.6 35B-A3B MoE (17.73 GB) or Qwen3.6 27B (16.82 GB) Gemma 4 E2B i1-Q6_K (3.85 GB) EmbeddingGemma 2 Q8_0 (0.31 GB) ~24.6 GB
Windows / Linux x64 (alpha) 50% RAM (+ GPU VRAM) Qwen3-ASR 0.6B int8 (0.88 GB) + pyannote/TitaNet (0.08 GB) Qwen3.5 4B Q4_K_M (2.74 GB) on Vulkan/CUDA or CPU Gemma 4 E2B (3.85 GB) or LFM2.5 1.2B (0.73 GB) EmbeddingGemma 2 Q8_0 (0.31 GB) ~4.7–7.9 GB

FAQ

Local AI memory and hardware questions

Can an 8 GB Apple Silicon Mac run local AI meeting notes and autocomplete?

Yes, with compact models. Pair Qwen3-ASR 0.6B (1.02 GB) or Parakeet TDT 0.6B v3 (0.60 GB) for transcription, Qwen3.5 0.8B (0.53 GB) for summaries, EmbeddingGemma 2 (0.31 GB) for search, and LFM2.5 1.2B (0.73 GB) for lightweight autocomplete. LokalBot's memory governor caps resident models at 50% of physical RAM (4 GB on an 8 GB Mac) and evicts idle models automatically when another task runs.

What is the best local AI model stack for a 16 GB Mac?

A 16 GB Apple Silicon Mac runs LokalBot's full default stack (~9.6 GB total on disk): Qwen3-ASR 1.7B (2.52 GB) for transcription, Nemotron 3 Core ML (0.20 GB) for speaker separation, Qwen3.5 4B Q4_K_M (2.74 GB) for summaries and work memory, EmbeddingGemma 2 Q8_0 (0.31 GB, 496 MB server peak) for hybrid search, and Gemma 4 E2B (3.85 GB) or LFM2.5 1.2B (0.73 GB) for autocomplete. ModelResidency automatically swaps idle models inside its 8 GB cap so macOS, browsers, and call apps keep half of system memory.

Why does total disk size differ from active RAM?

Two reasons: first, models do not all run at the same time — transcription and speaker diarization load after a meeting and unload when idle, before the summarizer runs. Second, an active language model needs extra memory beyond its GGUF file size for KV cache and runtime buffers (LokalBot budgets +3.0 GiB for the 32K-context main LLM, +768 MiB for 4K autocomplete, and +384 MiB for embeddings).

Do I need a dedicated GPU on Windows or Linux?

No. LokalBot's bundled llama.cpp runs on a Vulkan or CUDA GPU when available and falls back to x86_64 CPU otherwise. On an AMD Ryzen 7 9700X (CPU-only, 8 threads), Qwen3.5 4B finished short extraction in 2.86 seconds and a long-meeting stress test in 120 seconds with 6/6 native tool calls in our Linux CPU benchmark. An 8 GB+ GPU fits the default 4B summarizer and autocomplete models in VRAM for faster responses.

Does this calculator send my hardware specs to a server?

No. Every calculation runs locally in your browser with zero analytics scripts, cookies, or network requests. Share links encode only the selections you choose in the URL query string.

Try it locally

Run the whole local stack on your own machine.

Free, open source (GPLv3), and no account. Download only the models you pick.

Download for macOS

Apple Silicon · macOS 15+ · or with Homebrew: brew install --cask stevyhacker/tap/lokalbot

Benchmarks · System requirements · Meeting Migrator · All guides