_front.jpg){kind=link}
Best local AI models for a Mac Studio with 128 GB
A Mac Studio with 128 GB of unified memory has about 89.6 GB free for a model once macOS and your apps are running. These are the most downloaded open language and vision-language models that fit, for LM Studio, Ollama, llama.cpp or MLX.
128 GB unified memory · about 89.6 GB usable for a model · speeds estimated for M4 Max, 40-core GPU
Buy a Mac Studio on Amazon →
As an Amazon Associate we earn from qualifying purchases.
Last updated 3 November 2026 · refreshed after every Hugging Face sync · 15 models
| # | Model | Size | Memory | License | Downloads (30 days) | Likes | Est. speed (Q4) |
|---|---|---|---|---|---|---|---|
| 1 | Qwen3.8-27B Qwen | 27B · FP16 | ~65 GB | Apache 2.0 | 8M | 19.1K | ~25 tok/s |
| 2 | gemma-4-31B-it | 31B · FP16 | ~75 GB | Apache 2.0 | 8.1M | 4.3K | ~21 tok/s |
| 3 | gpt-oss-20b openai | 20B · FP16 | ~48 GB | Apache 2.0 | 6M | 5.3K | ~92 tok/s |
| 4 | Qwen3-14B Qwen | 14B · FP16 | ~34 GB | Apache 2.0 | 13M | 533 | ~42 tok/s |
| 5 | Qwen3.6-35B-A3B Qwen | 35B · FP16 | ~84 GB | Apache 2.0 | 4.6M | 3.1K | ~137 tok/s |
| 6 | gemma-4-26B-A4B-it | 26B · FP16 | ~63 GB | Apache 2.0 | 5.9M | 1.7K | ~91 tok/s |
| 7 | Qwen3.8-27B-FP8 Qwen | 27B · FP16 | ~65 GB | Apache 2.0 | 4.7M | 955 | ~25 tok/s |
| 8 | JEV-27B-VL autotrust | 27B · FP16 | ~65 GB | Apache 2.0 | 2.5M | 3.1K | – |
| 9 | Qwen3.6-35B-A3B-NVFP4 nvidia | 35B · FP16 | ~84 GB | Apache 2.0 | 3.8M | 674 | ~137 tok/s |
| 10 | Qwen3.6-35B-A3B-FP8 Qwen | 35B · FP16 | ~84 GB | Apache 2.0 | 4.1M | 440 | ~137 tok/s |
| 11 | Qwen3-32B Qwen | 32B · FP16 | ~77 GB | Apache 2.0 | 3M | 791 | ~20 tok/s |
| 12 | NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 nvidia | 30B · FP16 | ~72 GB | Custom (other) | 1.7M | 840 | – |
| 13 | dolphin-2.9.1-yi-1.5-34b dphn | 34B · FP16 | ~82 GB | Apache 2.0 | 4.8M | 90 | – |
| 14 | Qwen3-Coder-30B-A3B-Instruct Qwen | 30B · FP16 | ~72 GB | Apache 2.0 | 1.2M | 1.3K | ~104 tok/s |
| 15 | Qwen3-30B-A3B Qwen | 30B · FP16 | ~72 GB | Apache 2.0 | 1.1M | 982 | ~104 tok/s |
How we rank
Language and vision-language models in GGUF, MLX, FP16 format whose published weights need about 27–89.6 GB. We include models seen on Hugging Face in the last 14 days with at least 10,000 downloads in the last 30 days. Score = log10(30-day downloads) + ½ × log10(likes + 1), so popularity leads and likes break near-ties. We keep one entry per model name and leave out adult and uncensored builds. Memory is estimated from parameter count and the published file format (16-bit ≈ 2 bytes per parameter, 4-bit ≈ 0.5) plus 20% headroom; long contexts need more. A quantised copy of a bigger model may also fit — check the model page.
Speed data: ModelFit (CC BY 4.0), from its open dataset, refreshed weekly. These are modelled estimates of decode speed for a 4-bit (Q4) build of the same model at a 4K context, not measured benchmarks, and they are not an endorsement by ModelFit. A dash means ModelFit has no estimate for that model.