Can I run it on a Mac?

Pick a model and see which Macs run it. Each of the 122 Mac-runnable models in the LLMCheck index gets a verdict for all 54 Mac configurations: whether it fits macOS’s default GPU memory limit, the build, the speed with its provenance, and the context that fits. 35 fit the 16 GB Mac mini M6 inside its default limit.

Which Macs run this model?

Choose a model to list every Mac that runs it, grouped by verdict, with the build, speed and context limit for each. Every model also has its own page, listed below by lab.

Provenance key
sourced
Published figure — vendor or press review
estimated
community
Linked individual run — blog, forum or submission

Every model, by lab

Each page answers one question for one model: can a Mac run it, which one, how fast and with how much context. Listed with the memory it needs at a 4K context and how many of the 54 Mac configurations load it.

Alibaba

DeepSeek

Mistral

Google

Meta

Zhipu AI

NVIDIA

Ant Group

MiniMax

Liquid AI

Moonshot AI

Prism ML

Xiaomi

IBM

Microsoft

Nanbeige

OpenAI

OpenBMB

Poolside

StepFun

Swiss AI

Tencent

Thinking Machines

Aleph Alpha

Cohere

DeepGrove

HuggingFace

JetBrains

KRAFTON

Kwaipilot

MBZUAI IFM

Nous Research

Recursive

Reflection AI

Ricursive Intelligence

SparkLLM (XHToken)

Xiaohongshu (RedNote)

FAQ

How do I know if my Mac can run a local LLM?

Compare the model’s memory need with your Mac’s unified memory, then check macOS’s GPU limit and the context you need. Each model page here does that for all 54 Mac configurations in the LLMCheck index and gives a verdict, the build, the expected speed with its provenance, and the longest context that fits.

How much memory does macOS let the GPU use?

By default only part of it: on macOS 26 and later about 12 GB of a 16 GB Mac, about 25 GB of a 32 GB Mac, about 52 GB of a 64 GB Mac, about 108 GB of a 128 GB Mac; on macOS 15 and earlier less (11 of 16 GB, 21 of 32 GB, 48 of 64 GB, 96 of 128 GB). M5 and M6 Macs run macOS 26 or later. The limit is a performance threshold: past it, the runtime moves part of it to the CPU or shortens the context (llama.cpp, Ollama), runs it slowly (MLX), or stops with an out-of-memory error. sudo sysctl iogpu.wired_limit_mb raises the limit until the next restart; each model page gives the command for every Mac size that needs it.

Are these speeds measured?

Some are. Every speed carries its provenance: published (a vendor or a press review with its setup), community (one person’s linked run) or estimated (the memory-bandwidth model). 61 of the 122 Mac-runnable models have at least one measured run, listed on their page with its source. LLMCheck runs no tests of its own.

Which Mac runs the most models?

The Mac Studio M5 Ultra (512 GB) runs all 122 Mac-runnable models. Among smaller Macs, the Mac mini M6 (16 GB) runs 35 inside its default GPU limit, 1 more after raising it and 2 only on paper (over even a raised limit); the Mac mini M6 (32 GB) runs 63 inside its default GPU limit; the Mac mini M5 Pro (64 GB) runs 72 inside its default GPU limit, 2 more after raising it and 3 only on paper (over even a raised limit).

How is the context limit worked out?

From each model’s own config.json: the KV cache bytes per token of its attention layers (sliding-window and linear-attention layers counted as such), at the 16-bit default, added to the catalog’s minimum memory, which already covers a 4K-token prompt. A q8 KV cache halves the added part. These are estimates.

Cite this page

LLMCheck. (2026). Can I run it on a Mac? llmcheck.net. https://llmcheck.net/can-i-run/ (updated ). CC BY 4.0.

How to cite LLMCheck →Data: CC BY 4.0Markdown version

Related

Figures from the LLMCheck index, data version 2026-11-18. The picker reads /data/canirun.json (CC BY 4.0).