Best local LLMs for a 24 GB or 48 GB Mac (Nov 2026)

Short answer

On a 24 GB Mac, the safe picks are 7B to 9B models at Q8_0, about 10 GB: MiMo V2.6 9B at Q8_0 (9.55 GB) or Jais 3 7B at Q8_0 (9.57 GB). The 27B to 36B models fit only at 3 to 4 bits with a short context and little else open. On a 48 GB Mac, the 4-bit builds of Qwen 3.8 27B, KVARK Sparkle 27B, Jais 3 32B, Jais 3 MoVA 36B and MiniCPM-V 4.7 35B-A3B all fit with room for context.

The table ↓24 GB picks ↓48 GB picks ↓Model-fit calculator →

How I sized this list

macOS lets the GPU use about 75% of unified memory by default, which is about 18 GB on a 24 GB Mac and 36 GB on a 48 GB Mac. My own starting rule is stricter: keep the weights near half of memory, so the KV cache, the browser and macOS have room. The model-fit calculator does this sum properly for a chosen context length. Below, “fits” means the file is under the GPU limit with a few GB to spare, and “tight” means it is close enough that a long context or a busy machine will push it into swap.

I have not run these models. Sizes are read from the download pages on 4 November 2026, and every speed is labelled with who measured it.

What fits a 24 GB or 48 GB Mac

Recent open models, file size and fit (sizes are vendor or uploader file sizes)
Model	Build	File, GB	24 GB Mac	48 GB Mac
MiMo V2.6 Distill Qwen 9B	Q8_0 GGUF	9.55	Fits	Fits
Jais 3 7B	Q8_0 GGUF	9.57	Fits	Fits
Drex Max v1 27B	affine-q4 (MLX)	14.47	Fits, per Nace	Fits
MiniCPM-V 4.7 35B-A3B	mix_q3gud_im GGUF	15.5	Tight	Fits
KVARK Sparkle 1.0 27B	IQ4_XS GGUF	15.31	Tight	Fits
KVARK Sparkle 1.0 27B	Q4_K_M GGUF	16.81	Tight	Fits
Jais 3 32B	Q4_K_M GGUF	21.08	No	Fits
MiniCPM-V 4.7 35B-A3B	Q4_K_M + projector	22.3	No	Fits
Jais 3 MoVA 36B-A4B	Q4_K_M GGUF	22.37	No	Fits
KVARK Sparkle 1.0 27B	Q6_K GGUF	22.43	No	Fits
Drex Max v1 27B	Q8_0 GGUF	27.25	No	Fits, per Nace
KVARK Sparkle 1.0 27B	Q8_0 GGUF	29.05	No	Tight
Jais 3 32B	Q8_0 GGUF	36.97	No	No

Best picks for a 24 GB Mac

General chat and coding help: MiMo V2.6 9B. A Qwen 3.5 9B tuned on Xiaomi’s larger models. One community run measured 24 tokens a second at Q8_0 on an M4 Pro. My MiMo note has the sizes and Ollama tags.
Arabic: Jais 3 7B. IFM’s own Q8_0 file is 9.57 GB, and the Q4_K_M is 5.59 GB. On IFM’s card it beats Qwen3.5-9B on Arabic MMLU, 85.2 to 80.9. See running Jais 3 locally.
Images and documents: MiniCPM-V 4.7, carefully. The 15.5 GB file runs at about 25 tokens a second when it stays in memory, but Micdrop saw two-minute requests once a browser and editor pushed it into swap. The MiniCPM note covers the 16 GB route too.

A 27B at 4-bit will load on 24 GB, but at that point the context and every other app are competing for the last few GB. If 27B quality matters to you, 48 GB is the better buy.

Best picks for a 48 GB Mac

Coding and agents: Qwen 3.8 27B. Still the reference model in this size. My Qwen 3.8 note has the measured runs on my own M4 Max.
Shorter reasoning: KVARK Sparkle 1.0 27B. Egzakta kept Qwen 3.8 27B’s architecture and trained it to think less. On Egzakta’s numbers it uses 34 to 67% fewer reasoning tokens at about the same accuracy (LiveCodeBench 82.7 against 77.3 for the base), and its median time per answer on an H200 was 13.7 seconds against 28.1. Apache 2.0. The Q4_K_M is 16.81 GB and the Q6_K 22.43 GB. Nobody I found has published a Mac run yet.
Arabic: Jais 3 MoVA 36B-A4B or 32B. Both are about 22 GB at Q4_K_M. MoVA uses about 4B parameters per token, so it should be faster.
Vision: MiniCPM-V 4.7 35B-A3B. Collected oMLX runs (secondary) show 75 to 82 tokens a second at 4-bit on a 48 GB M4 Pro, with about 21 GB peak.
Faster MoE alternatives: Nemotron 3.5 Lightning has 18 to 25 GB files and runs about 100 tokens a second on an M4 Max. Nanbeige 4.3 33B is the other recent dense option, compared directly against Qwen.

A note on Drex Max v1

Nace.AI’s Drex Max v1 is a dense 27B that turns up in “new models” lists. It is a decision model with a fixed set of 255 outputs, so it classifies inputs and does not write text. Nace says the affine-q4 build (14.47 GB) runs on 24 GB Macs through its Kapitsa MLX runtime, and the Q8_0 GGUF (27.25 GB) needs 48 GB. The llama.cpp and Ollama paths use Nace’s own forks, and the license is Open RAIL-M rather than Apache. Only download it if you have a classification job for it.

FAQ: local models on 24 GB and 48 GB Macs

What is the best local LLM for a 24 GB Mac in November 2026?

For most people, a 9B-class model at Q8_0, such as MiMo V2.6 9B (9.55 GB). It leaves room for a long context. For Arabic, Jais 3 7B at the same size.

Can a 24 GB Mac run a 27B model?

At 3 to 4 bits, yes, with a short context. KVARK Sparkle’s IQ4_XS file is 15.31 GB. Expect swap if a browser and editor are also open.

What can a 48 GB Mac run that a 24 GB one cannot?

The 4-bit builds of 27B to 36B models with real context: Qwen 3.8 27B, KVARK Sparkle, Jais 3 32B and MoVA, and MiniCPM-V 4.7 35B-A3B.

Is 48 GB enough for Q8_0 27B models?

Just. KVARK Sparkle at Q8_0 is 29.05 GB, under the default 36 GB GPU limit but with little room for context.

Should I buy 24 GB or 48 GB for local AI?

If you want 27B-class models as daily tools, 48 GB. On 24 GB you will mostly run 7B to 9B models well and larger ones badly.

Related local notes

Sources and what I did not test

I have not run any model in this table. Fit labels are my planning estimates from file sizes, not measurements.

Sources

Suhail Mohebi · published 4 November 2026 · the cover is a generated illustration