Apple Silicon Mac configurations
Every Apple Silicon chip from M1 to M6 in every memory size Apple sells it with: unified memory, memory bandwidth, GPU core options and the Macs it ships in, read from Apple's Tech Specs pages, with the source of every row. On the site: every Mac.
- JSON · 46 KB
76320724a7fac765… - CSV · 26 KB
e8a976f183d2a114…
14 columns · 58 rows · checked 2026-10-19
| Column | Type | Meaning |
|---|---|---|
mac | string | Identifier: chip and memory size, as in mlx.app/mac/<mac> |
chip | string | Chip name |
chip_id | string | Chip identifier |
generation | number | M-series generation (1 = M1) |
tier | string | base, pro, max or ultra |
memory_gb | number | Unified memory, GB |
bandwidth_gbs | number | Memory bandwidth Apple publishes, GB/s (empty where Apple publishes none: M1) |
bandwidth_gbs_used | number | Bandwidth the estimates use, GB/s (M1: 68, from Apple's relative statements; see note) |
gpu_memory_default_gb | number | About 75% of unified memory: what macOS lets the GPU use by default (derived) |
gpu_cores | string | GPU core options with this memory size (list) |
macs | string | Macs the chip ships in, by Apple's model names (list) |
note | string | Caveats from the sources |
page | string | Page on mlx.app with every model this Mac runs |
sources | string | Apple pages the figures were read from (list) |
Open AI models: parameters, active parameters and memory
Every open-weight language model on mlx.app/compare: total and active parameters, architecture (dense or mixture of experts), context length, licence, release date, precision and weight size, read from each model's official Hugging Face card, config.json and file listing. On the site: every model.
- JSON · 57 KB
bd3695ec1905374d… - CSV · 33 KB
99ebd96d8148ef16…
19 columns · 82 rows · checked 2026-10-19
| Column | Type | Meaning |
|---|---|---|
model | string | Identifier, as in mlx.app/run/<model> |
name | string | Model name |
maker | string | Who released it |
hf_repo | string | Official Hugging Face repository |
params_b | number | Total parameters, billions (model card) |
active_params_b | number | Parameters read per generated token, billions (equals params_b for dense models, except a looped model, which reads its shared block once per loop: at its default loop count) |
architecture | string | dense or moe |
context_tokens | number | Maximum context, tokens (config.json) |
license | string | Licence named on the card |
released | string | Release date (YYYY-MM-DD; the repository's creation date where the card gives none) |
modality | string | Inputs: text, or text+image |
precision | string | Precision of the official weights |
repo_weights_gb | number | Size of the official weight files, decimal GB |
native_gb | number | Weights in the model's native low-bit format, GB (only models released that way) |
native_bits | number | Bits per weight of that native format |
offload_gb | number | Lookup tables a runtime can leave on disk, GB (not counted in memory estimates) |
mlx_build | string | A community MLX build on Hugging Face, where one exists |
page | string | Page on mlx.app with the memory it needs and every Mac that runs it |
sources | string | Pages the figures were read from (list) |
Which open model runs on which Mac, and how fast (estimates)
Every Apple Silicon Mac configuration against every open model: the highest-precision version that fits in the memory macOS gives the GPU, how much memory it needs with room for context, and the estimated writing speed range in tokens a second. Estimates from published specs, not measurements; the method is at mlx.app/compare. On the site: the calculator.
- JSON · 1.5 MB
10aabea1c1684fcd… - CSV · 525 KB
cc9d62f957171a0f…
16 columns · 4,698 rows · checked 2026-10-19
| Column | Type | Meaning |
|---|---|---|
mac | string | Mac configuration (see macs) |
chip | string | Chip name |
memory_gb | number | Unified memory, GB |
bandwidth_gbs | number | Memory bandwidth used for the estimate, GB/s |
model | string | Model (see models) |
model_name | string | Model name |
version | string | Version recommended: the highest precision that fits (8-, 6-, 4-, 3- or 2-bit MLX, or the native format) |
bits_per_weight | number | Bits stored per weight, including MLX's group scales (+0.5) |
weights_gb | number | Weights in memory, GB |
needs_gb | number | Weights plus room for context and runtime (8% or 1.5 GB, whichever is more), GB |
fit | string | fits (within the GPU's default 75%), tight (within 90% of memory), no |
group | string | Runs well (fits, 10+ tok/s), Runs, but slowly, Only just fits, Too big |
below_4_bit | boolean | True when only a version under 4 bits fits (quality drops) |
tok_s_low | number | Estimated writing speed, low end (45% of the bandwidth ceiling), tokens/s; empty when it doesn't fit |
tok_s_high | number | Estimated writing speed, high end (80% of the ceiling), tokens/s |
tok_s_ceiling | number | Bandwidth ceiling: bandwidth / bytes read per token, tokens/s |
Estimates, and how they’re worked out
The Mac and model tables are facts copied from Apple’s Tech Specs and each model’s official Hugging Face card; every row names its source. The fit table is worked out, not measured: a model fits when its weights at a given precision (bits per weight plus MLX’s half-bit group scales), with room for context, stay within the 75% of unified memory macOS lets the GPU use by default. Writing speed is memory bandwidth divided by the bytes read per token, shown as 45% to 80% of that ceiling, the range real runs reach in llama.cpp’s public Apple Silicon benchmarks. Prompt reading, long contexts and batching are not covered. The full method: mlx.app/compare. Measure your own Mac on MLX Bench.
Not included: Artificial Analysis Intelligence Index scores (shown on the pages with a link, but their terms reserve redistribution) and llama.cpp benchmark results (linked from the method, not copied).
Use it, and cite it
Free under Creative Commons Attribution 4.0: copy, change and publish it, commercially too, with a link to this page. The specifications themselves are facts from Apple and the model makers; model and product names belong to their owners. MLX is not affiliated with Apple Inc.
MLX Network (2026). MLX open data: Apple Silicon Macs, open AI models and what runs where (version 2026-10-19). https://mlx.app/data. CC BY 4.0.The URLs above always serve the latest version; each file carries its version (version in JSON, the X-Dataset-Version header, and the file name it saves as). index.json lists every file with its size and SHA-256.
More open data
- Explorer API: every block, transaction, account and seat on the MLX Network, as JSON.
- llms.txt and llms-full.txt: the site in plain text for AI assistants, with every guide in full.
- Sitemap: every page, with the date its facts were last checked.