Data sources, method and disclosure

Affiliate and sponsor disclosure

How "fits" is calculated

• Weights: Weights = size of the public GGUF file (Q4_K_M or Q8_0) on Hugging Face, in bytes. Not estimated from parameter count.
• KV cache: fp16 (2 bytes/element) KV cache; K and V both counted (Gemma 4 global layers conservatively counted as separate K and V even though config says K=V). KV(ctx) = full_bytes_per_token*ctx + swa_bytes_per_token*min(ctx,window) + fixed_state_bytes; bytes derived from config.json (layers, kv heads, head_dim, layer_types).
• Overhead: Flat 1 GiB for compute buffers/runtime. ASSUMPTION; real overhead varies with engine and batch size.
• Memory limit: Apple "GB" of unified memory is treated as GiB (2^30 bytes). Model file sizes (decimal GB on Hugging Face) are converted to GiB. https://tools.its-tahir.com/guides/run-llm-locally-on-mac (about 2/3 of unified memory on Macs with <=32GB, 3/4 on larger); llama.cpp maintainer says ~75% (https://github.com/ggml-org/llama.cpp/discussions/4167). Default Metal limit can be raised with sysctl iogpu.wired_limit_mb.
• Headroom: "Fits with tuning" assumes you raise the GPU wired limit and leave ~4 GiB for macOS and apps. ASSUMPTION, not measured.
• Statuses: fits: need <= 85% of default GPU limit; tight: need <= default GPU limit; tuned: need <= memory - headroom (requires raising the GPU wired limit); no: otherwise. need = weights + KV + overhead.
• Tight margin: "Tight" = fits under the default GPU limit but uses >85% of it. Margin is an ASSUMPTION: in one public run (stared/benching-local-llms-on-apple-silicon) llama.cpp process RSS was ~44 GB for a 36.9 GB Q8_0 GGUF of Qwen3.6-35B-A3B, i.e. real usage can exceed file size + KV by several GB.
• Bandwidth ceiling (optional toggle): Bandwidth ceiling tok/s = memory_bandwidth_GB/s / (bytes read per token in GB). Bytes read per token = Q4 file size * (active params / total params) for MoE; = Q4 file size for dense. An upper bound for decode, NOT a measurement; real speeds are typically well below it.

Speed measurements (tokens/sec)

Shown only where a public source exists. Different engines, quantizations, contexts and llama.cpp builds are not directly comparable; each row says which it is. A blank means nobody has published a measurement we could find, not that the Mac is slow.

Macs in the database

Prices are US list prices from apple.com on 2026-10-16; configurations with a memory upgrade use apple.com's base price plus the configurator delta reported publicly (see basis). "n/a" = not captured; we do not guess.

Models in the database

Known gaps

• Most Mac × model combinations have no public speed measurement.
• Memory-upgrade prices for MacBook Air, MacBook Pro and iMac were not found in a public source and are left blank.
• Some Mac variants (M5 Max 32-core GPU, M5 Ultra 64-core GPU, M6 16GB, A18 Pro) have no public llama.cpp reference run.
• MLX-native quantizations are not modeled; sizes are GGUF.
• DeepSeek V4 Flash and other very large MoE models are omitted because their KV cache cannot be derived reliably from config.json.

Trademarks

Mac, MacBook, MacBook Air, MacBook Pro, iMac, Mac mini, Mac Studio and Apple silicon are trademarks of Apple Inc. This site is independent and not endorsed by Apple.