Last updated Oct 2026

Which open-weight model can my Mac run?

A plain-English directory of today’s major open-weight model families, how big they are, roughly how much memory they need at 4-bit, and whether one Mac can run them.

How to read this: “Total” is the full parameter count, which decides memory. “Active” is how many parameters work on each word in mixture-of-experts (MoE) models, which decides speed. “~4-bit size” is an estimate of the model file at 4-bit (about 0.55–0.6GB per billion parameters). You need extra memory on top for macOS and for the conversation itself. What do these terms mean?

Family	Model	Total params	Active	~4-bit size	Suggested Mac memory	Single Mac?	Notes
Liquid LFM2.5	LFM2.5 1.2B / 2.6B	1.2B / 2.6B	all	~1–2GB	Any (8GB+)	Yes	Tiny, fast, built for on-device agents
Liquid LFM2.5	LFM2.5-8B-A1B	8B	1.5B	~5GB	16GB	Yes	MoE; very fast
Liquid LFM2.5	LFM2.5-24B-A2B (Oct 2026)	24B	2B	~14GB	24GB	Yes	Largest LFM2.5; reasoning model
Google Gemma 4	Gemma 4 E4B / 12B	8B / 12B	all	~5 / ~7GB	16GB	Yes	Strong small all-rounders
Google Gemma 4	Gemma 4 26B A4B	25.2B	3.8B	~15GB	24GB	Yes	MoE; fast for its quality
Google Gemma 4	Gemma 4 31B	~31B	all	~18GB	32GB	Yes	Dense; slower than the 26B MoE
Alibaba Qwen	Qwen3.5 9B	9.7B	all	~6GB	16GB	Yes
Alibaba Qwen	Qwen3.8 27B (Aug 2026)	27.8B	all	~16GB	24–32GB	Yes	Popular mid-size pick
Alibaba Qwen	Qwen3.6 35B-A3B	~35B	3B	~20GB	32GB	Yes	MoE; quick on Pro/Max chips
Alibaba Qwen	Qwen3.5 122B-A10B	122B	10B	~70GB	96–128GB	Yes
Alibaba Qwen	Qwen3.8-Flash-Next (preview, Aug 2026)	125B backbone + extra embeddings	~6B	varies; compressed builds reported to fit 128GB	128GB	Yes	Experimental preview
Alibaba Qwen	Qwen3.8 2.4T-A95B	~2.4T	95B	1TB+	—	No	Datacenter only
Mistral	Ministral 3 14B	13.5B	all	~8GB	16GB (tight) / 24GB	Yes
Mistral	Mistral Small 3.2 24B	24B	all	~14GB	24–32GB	Yes
Mistral	Mistral Small 4	119B	6.5B	~67GB	96–128GB	Yes	MoE
Mistral	Mistral Medium 3.5	128B	all	~72GB	96–128GB	Yes	Dense, so slow on a Mac
Meta Llama	Llama 4 Scout	109B	17B	~61GB	96–128GB	Yes	Released Apr 2025
Meta Llama	Llama 4 Maverick	400B	17B	~225GB	256GB+ (tight)	Big Mac only
Meta Muse	Muse Spark Open (Oct 2026)	816B	42B	~460GB	Two M5 Ultra Mac Studios	No	Meta says it’s optimized for two M5 Ultra Mac Studios with MLX
Zhipu GLM	GLM-5.3-Flash (Aug 2026)	~320B	18B	~180GB	256GB	Big Mac only
Zhipu GLM	GLM-5.3 (Aug 2026)	~744B	40B	~420GB	512GB (very tight) or 2 Macs	Big Mac only	Flagship
DeepSeek	DeepSeek V4-Flash	284B	13B	~160GB	256GB	Big Mac only
DeepSeek	DeepSeek V4.1-Flash (Sep 2026)	~552B + 196B memory module	8–16B	~420GB+	512GB at best, or 2 Macs	Big Mac only	Most people use DeepSeek’s app instead
DeepSeek	DeepSeek V4-Pro	~1.6T	49B	~900GB	—	No	Datacenter only

Sizes are estimates derived from published parameter counts; real file sizes differ by quantization method. Model releases move fast, so check the model card on Hugging Face before downloading. “Big Mac only” means a 256–512GB Mac Studio, and even then it may be tight.

Our picks by Mac

16GB

Gemma 4 12B or Qwen3.5 9B for everyday chat; LFM2.5 models for speed.

24–32GB

Qwen3.8 27B or Gemma 4 26B A4B. At 32GB, try Qwen3.6 35B-A3B for speed.

64GB

Run the 27–35B models at 8-bit for better quality and long documents.

128GB

Qwen3.5 122B-A10B, Mistral Small 4 or a compressed Qwen3.8-Flash-Next.

Families at a glance

Qwen (Alibaba): The broadest range of sizes, from small to trillion-scale. Most models use the permissive Apache 2.0 license; some newer ones use Qwen’s own licenses.
Gemma 4 (Google DeepMind): Small and mid-size models (about 5B–31B) under Apache 2.0. A natural fit for 16–32GB Macs.
GLM (Zhipu / Z.ai): Large MoE models known for coding and agent work. The Flash version is the more Mac-friendly one.
DeepSeek: Very large MoE models with an MIT license. Mostly too big for a single Mac; most people use DeepSeek’s app.
Llama 4 (Meta): Scout and Maverick, both MoE with 17B active, released in April 2025.
Muse Spark Open (Meta): Released October 15, 2026: 816B total, 42B active, multimodal, under Meta’s Muse Model License. Not a single-Mac model.
Mistral: From compact Ministral models to the 119B Mistral Small 4 MoE and the 128B dense Medium 3.5.
Liquid LFM2.5 (Liquid AI): Small, fast models (230M–24B) designed for on-device use, with GGUF and MLX builds published by Liquid.

Licenses vary and some restrict commercial use. Read the license on the model card.