Mac M2 Max 96G — what LLMs can it run?
97 of 110 indexed models fit comfortably in 96GB at 4K context, each at the highest-quality quant that still leaves headroom.
The short answer for Mac M2 Max 96G
96 GB, 400 GB/s Unified LPDDR5. 97 of 110 models in this index fit comfortably at 4K context; 99 load at all.
- Biggest that fits
- GPT-OSS 120B — 62.7 GB at Q4_K_M — 9.3 GB spare at 4K, so longer context comes out of a thin margin.
- Room to grow
- Nemotron 3.5 Bolt 74B-A9B — 23.1 GB at IQ2_XXS — under 60% of the card, which leaves 49.0 GB for a long context window or a second process.
- Speed ceiling
- Llama 3.1 8B Instruct at Q4_K_M reads 4.6 GB of weights per token, so 400 GB/s puts a hard ceiling near 87 tok/s. That is arithmetic on two published numbers, not a benchmark — real throughput lands below it. No run on this card has been measured here, so there is nothing to compare the ceiling against.
- The wall you will hit
- The GPU's share of unified memory, not the 96 GB on the box. macOS reserves part of the pool for the system and caps what a single process may wire down, so the practical budget is meaningfully below nameplate — the limit is adjustable (`iogpu.wired_limit_mb`) but it is not absent. Bandwidth is 400 GB/s, which is the number that decides tok/s once a model fits.
70B+ · 10
| Model | Quant | Est. VRAM | Headroom | tok/s |
|---|---|---|---|---|
| GPT-OSS 120B117B MoE | Q4_K_M | 62.73 GB | +9.3 GB | ≤161Ceiling |
| Llama 4 Scout 17B (16E)109B MoE | Q3_K_M | 55.92 GB | +16.1 GB | ≤51Ceiling |
| GLM-4.5-Air106B MoE | Q3_K_M | 54.37 GB | +17.6 GB | ≤73Ceiling |
| Llama 3.2 90B Vision Instruct90B | Q4_K_M | 57.5 GB | +14.5 GB | ≤8Ceiling |
| Nemotron 3.5 Bolt 74B-A9B74B-A9B | IQ2_XXS | 23.05 GB | +49 GB | ≤157Ceiling |
| Qwen2.5 72B Instruct72B | Q5_K_M | 55.31 GB | +16.7 GB | ≤8Ceiling |
| Llama 3.1 70B Instruct70B | Q5_K_M | 53.75 GB | +18.3 GB | ≤8Ceiling |
| Llama 3.3 70B Instruct70B | Q5_K_M | 53.75 GB | +18.3 GB | ≤8Ceiling |
| DeepSeek-R1-Distill-Llama-70B70B | Q4_K_M | 46.1 GB | +25.9 GB | ≤10Ceiling |
| Jamba 1.5 Mini52B-A12B | Q4_K_M | 33.01 GB | +39 GB | ≤58Ceiling |
32B · 30
| Model | Quant | Est. VRAM | Headroom | tok/s |
|---|---|---|---|---|
| Kimi Linear 48B-A3B Instruct48B-A3B | Q8_0 | 53.33 GB | +18.7 GB | ≤132Ceiling |
| Mixtral 8x7B Instruct47B MoE | Q4_K_M | 30.13 GB | +41.9 GB | ≤54Ceiling |
| Seed-OSS 36B Instruct36B | Q5_K_M | 27.81 GB | +44.2 GB | ≤16Ceiling |
| Qwen3.6 35B-A3B35B-A3B | Q8_0 | 39.6 GB | +32.4 GB | ≤132Ceiling |
| Command R 35B35B | Q4_K_M | 22.86 GB | +49.1 GB | ≤20Ceiling |
| Yi 1.5 34B Chat34B | Q4_K_M | 22.82 GB | +49.2 GB | ≤20Ceiling |
| Qwen3 32B Instruct32B | Q4_K_M | 21.85 GB | +50.2 GB | 12–15Derived |
| Qwen2.5 32B Instruct32B | Q4_K_M | 21.69 GB | +50.3 GB | 12–16Derived |
| Qwen2.5-Coder 32B Instruct32B | Q4_K_M | 21.69 GB | +50.3 GB | 12–16Derived |
| DeepSeek-R1-Distill-Qwen-32B32B | Q4_K_M | 21.69 GB | +50.3 GB | 12–16Derived |
| Nemotron 3.5 Lightning 30B-A3B30B-A3B | Q8_0 | 35.14 GB | +36.9 GB | ≤132Ceiling |
| Gemma 4 31B IT31B | Q8_0 | 35.72 GB | +36.3 GB | 7–9Derived |
| Qwen3 30B-A3B Instruct30B-A3B | Q5_K_M | 23.04 GB | +49 GB | ≤198Ceiling |
| Qwen3-Coder 30B-A3B Instruct30B-A3B | Q5_K_M | 23.04 GB | +49 GB | ≤198Ceiling |
| Qwen3-VL 30B-A3B Instruct30B-A3B | Q8_0 | 34.28 GB | +37.7 GB | ≤132Ceiling |
| GLM-4.7-Flash30B-A3B | Q8_0 | 33.47 GB | +38.5 GB | ≤132Ceiling |
| Granite 4.2 30B30B | Q8_0 | 33.61 GB | +38.4 GB | 7–10Derived |
| Muse Glimmer 30B30B | Q8_0 | 31.08 GB | +40.9 GB | 8–10Derived |
| Qwen3.8 27B27B | Q4_K_M | 17.47 GB | +54.5 GB | 14–19Derived |
| Qwen3.6 27B27B | Q8_0 | 30.71 GB | +41.3 GB | 8–10Derived |
| Gemma 3 27B IT27B | Q4_K_M | 18.37 GB | +53.6 GB | 14–18Derived |
| Gemma 2 27B Instruct27B | Q5_K_M | 21.76 GB | +50.2 GB | 12–16Derived |
| Gemma 4 26B-A4B IT26B-A4B | Q8_0 | 28.42 GB | +43.6 GB | ≤99Ceiling |
| Mistral Small 24B Instruct24B | Q4_K_M | 15.89 GB | +56.1 GB | 16–21Derived |
| LFM2.5 24B-A2B24B-A2B | Q8_0 | 26.73 GB | +45.3 GB | ≤162Ceiling |
| Devstral Small 1.1 24B24B | Q6_K | 20.91 GB | +51.1 GB | 12–16Derived |
| Magistral Small 1.2 24B24B | Q6_K | 20.91 GB | +51.1 GB | 12–16Derived |
| Codestral 22B22B | Q4_K_M | 15.03 GB | +57 GB | 17–23Derived |
| ERNIE 4.5 21B-A3B21B-A3B | Q8_0 | 24.44 GB | +47.6 GB | ≤132Ceiling |
| GPT-OSS 20B21B MoE | MXFP4 | 12.76 GB | +59.2 GB | ≤202Ceiling |
14B · 18
| Model | Quant | Est. VRAM | Headroom | tok/s |
|---|---|---|---|---|
| InternLM2 20B Chat20B | Q5_K_M | 15.59 GB | +56.4 GB | 16–22Derived |
| DeepSeek-Coder-V2-Lite Instruct16B-A2.4B | Q8_0 | 17.56 GB | +54.4 GB | ≤165Ceiling |
| DeepSeek-V2-Lite Chat16B-A2.4B | Q4_K_M | 10.08 GB | +61.9 GB | ≤289Ceiling |
| StarCoder2 15B15B | Q4_K_M | 10.19 GB | +61.8 GB | 25–34Derived |
| Qwen3 14B Instruct14B | Q5_K_M | 11.65 GB | +60.4 GB | 22–29Derived |
| Qwen2.5 14B Instruct14B | Q5_K_M | 11.73 GB | +60.3 GB | 22–29Derived |
| DeepSeek-R1-Distill-Qwen-14B14B | Q4_K_M | 10.14 GB | +61.9 GB | 26–34Derived |
| Phi-4 14B14B | Q5_K_M | 11.77 GB | +60.2 GB | 22–29Derived |
| Phi-3 Medium 14B Instruct14B | Q6_K | 12.86 GB | +59.1 GB | 20–27Derived |
| Mistral Nemo 12B Instruct12B | Q6_K | 11.14 GB | +60.9 GB | 23–31Derived |
| Gemma 3 12B IT12B | Q5_K_M | 9.84 GB | +62.2 GB | 27–35Derived |
| Mellum2.1 12B-A2.5B Thinking12B-A2.5B | Q8_0 | 13.6 GB | +58.4 GB | ≤159Ceiling |
| Stable LM 2 12B Chat12B | Q4_K_M | 8.35 GB | +63.7 GB | 31–42Derived |
| Llama 3.2 11B Vision Instruct11B | Q8_0 | 12.9 GB | +59.1 GB | 20–26Derived |
| Solar 10.7B Instruct11B | Q4_K_M | 7.6 GB | +64.4 GB | 36–47Derived |
| Falcon 3 10B Instruct10B | Q4_K_M | 7.28 GB | +64.7 GB | 37–48Derived |
| Gemma 2 9B Instruct9B | Q8_0 | 11.7 GB | +60.3 GB | 24–31Derived |
| Qwen3.5 9B9B | Q8_0 | 10.37 GB | +61.6 GB | 24–31Derived |
7B · 25
| Model | Quant | Est. VRAM | Headroom | tok/s |
|---|---|---|---|---|
| GLM-4-9B-Chat9B | Q8_0 | 10.16 GB | +61.8 GB | 24–32Derived |
| Qwen3-VL 8B Instruct8B | Q8_0 | 10.39 GB | +61.6 GB | 25–33Derived |
| Granite 4.2 8B8B | Q8_0 | 10.45 GB | +61.6 GB | 25–33Derived |
| Ministral 3 8B Instruct8B | Q8_0 | 10.02 GB | +62 GB | 26–34Derived |
| Qwen2-VL 7B Instruct7B | Q4_K_M | 5.49 GB | +66.5 GB | 46–61Derived |
| Granite 3.1 8B Instruct8B | Q4_K_M | 5.88 GB | +66.1 GB | 46–62Derived |
| Qwen3 8B Instruct8B | Q6_K | 7.64 GB | +64.4 GB | 34–46Derived |
| Llama 3.1 8B Instruct8B | Q8_0 | 9.47 GB | +62.5 GB | 27–36Derived |
| Nous Hermes 3 Llama 3.1 8B8B | Q4_K_M | 5.64 GB | +66.4 GB | 47–63Derived |
| Aya 23 8B8B | Q4_K_M | 5.64 GB | +66.4 GB | 47–63Derived |
| OpenChat 3.6 8B8B | Q4_K_M | 5.64 GB | +66.4 GB | 47–63Derived |
| DeepSeek-R1-Distill-Llama-8B8B | Q5_K_M | 6.51 GB | +65.5 GB | 40–54Derived |
| InternLM2 7B Chat7B | Q4_K_M | 5.45 GB | +66.6 GB | 49–65Derived |
| Qwen2.5 7B Instruct7B | Q6_K | 6.77 GB | +65.2 GB | 37–49Derived |
| Qwen2.5-Coder 7B Instruct7B | Q4_K_M | 5.07 GB | +66.9 GB | 50–66Derived |
| DeepSeek-R1-Distill-Qwen-7B7B | Q4_K_M | 5.07 GB | +66.9 GB | 50–66Derived |
| Gemma 4 E4B ITE4B | Q8_0 | 8.46 GB | +63.5 GB | 29–38Derived |
| OLMo 2 7B Instruct7B | Q8_0 | 10.3 GB | +61.7 GB | 30–39Derived |
| OLMo 3 7B Instruct7B | Q8_0 | 10.3 GB | +61.7 GB | 30–39Derived |
| WizardLM-2 7B7B | Q4_K_M | 5.14 GB | +66.9 GB | 53–70Derived |
| Mistral 7B Instruct v0.37B | Q6_K | 6.75 GB | +65.3 GB | 39–51Derived |
| Zephyr 7B Beta7B | Q6_K | 6.75 GB | +65.3 GB | 39–51Derived |
| Gemma 4 E2B ITE2B | Q8_0 | 5.2 GB | +66.8 GB | ≤85Ceiling |
| Qwen3.5 4B4B | Q8_0 | 4.97 GB | +67 GB | ≤91Ceiling |
| Gemma 3 4B IT4B | Q8_0 | 5.05 GB | +67 GB | ≤92Ceiling |
≤3B · 14
| Model | Quant | Est. VRAM | Headroom | tok/s |
|---|---|---|---|---|
| Qwen3 4B Instruct4B | Q6_K | 4.05 GB | +68 GB | ≤128Ceiling |
| Phi-4 Mini Instruct3.8B | Q8_0 | 4.81 GB | +67.2 GB | ≤103Ceiling |
| Phi-3.5 Mini Instruct3.8B | Q8_0 | 5.89 GB | +66.1 GB | ≤104Ceiling |
| Granite 4.2 3B3B | Q8_0 | 4.41 GB | +67.6 GB | ≤108Ceiling |
| Llama 3.2 3B Instruct3B | Q8_0 | 4.05 GB | +68 GB | ≤123Ceiling |
| Qwen2.5 3B Instruct3B | Q8_0 | 3.59 GB | +68.4 GB | ≤128Ceiling |
| SmolLM3 3B3B | Q8_0 | 3.73 GB | +68.3 GB | ≤129Ceiling |
| Gemma 2 2B Instruct2B | Q8_0 | 3.34 GB | +68.7 GB | ≤152Ceiling |
| Qwen3.5 2B2B | Q8_0 | 2.23 GB | +69.8 GB | ≤202Ceiling |
| Qwen3 1.7B Instruct1.7B | Q8_0 | 2.39 GB | +69.6 GB | ≤230Ceiling |
| Qwen2.5 1.5B Instruct1.5B | Q8_0 | 1.83 GB | +70.2 GB | ≤258Ceiling |
| Llama 3.2 1B Instruct1B | Q8_0 | 1.51 GB | +70.5 GB | ≤320Ceiling |
| Qwen3.5 0.8B0.8B | Q8_0 | 0.94 GB | +71.1 GB | ≤494Ceiling |
| Qwen2.5 0.5B Instruct0.5B | Q8_0 | 0.6 GB | +71.4 GB | ≤816Ceiling |
How this list is built
Each row is the lowest-perplexity-loss quant of that model whose estimated total — weights plus KV cache at 4K context plus activation buffer — uses at most 88% of the card. That is the calculator's "green" threshold, so every row here has real headroom rather than only just fitting. Raise the context length and the list shortens; the calculator lets you check any combination directly. The speed column is this card: a measured run, a range scaled from this card's own Llama 2 7B benchmark, or a bandwidth ceiling. A figure that only exists for the RTX 4090 is labelled as that card.
A further 2 models load but with no headroom to spare (up to 105% of VRAM) — the Quant Hub’s GPU chips count those too, which is why its number is higher.
Measured on this card
No benchmark runs in this index were recorded on a Mac M2 Max 96G. Every figure on this page is calculated from the model architecture and the quant level — treat them as estimates, not measurements.
Setup guides for this card
Step-by-step guides written for the Mac M2 Max 96G's runtime or memory size, with the commands and the checks that prove the model is running on the card.
Stepping up
A Mac M5 Max 128G (128GB) fits 3 more of the indexed models than this card. Mac M5 Max 128G →
Common questions
What is the best local LLM for a Mac M2 Max 96G?
For everyday use, Nemotron 3.5 Bolt 74B-A9B at IQ2_XXS — about 23.1 GB of the card's 96 GB at 4K context, leaving 49.0 GB for a longer window. If you want the largest thing that will load, that is GPT-OSS 120B at Q4_K_M (62.7 GB, 9.3 GB spare). "Best" here means best fit for the memory budget — this index does not run task benchmarks, so it cannot tell you which model is smarter.
Can a Mac M2 Max 96G run Nemotron 3.5 Super VL 120B-A12B?
Only just. Its smallest build here, Q4_0, needs about 75.1 GB against 72.0 GB usable — that loads with nothing else running, with no margin for a longer context window. It is not on the list above, which requires a model to stay inside 88% of usable capacity.
How many tokens per second does a Mac M2 Max 96G do on an 8B model at Q4?
This index has no measured run on this card, so it does not publish a figure. What can be stated from specifications: generating a token requires reading every weight once, Llama 3.1 8B Instruct at Q4_K_M is 4.6 GB of weights, and this card moves 400 GB/s — a ceiling near 87 tok/s. Batch size, context length, the runtime and how much of the model sits in cache all take you below it.
Mac M2 Max 96G or Mac M5 Max 128G for local LLMs?
They do not hold the same set: Mac M2 Max 96G fits 97 and Mac M5 Max 128G 100 of 110 at 4K — macOS keeps part of the Mac M2 Max 96G's unified memory from the GPU, leaving about 72 GB usable. Speed is the other difference — 400 GB/s against 614 GB/s, a 1.5× gap in how fast the weights can be read, which is what token generation is bound by. The Mac M5 Max 128G generates faster on any model both can hold.
97 of 110 indexed models fit comfortably in 96GB at 4K context, each at the highest-quality quant that still leaves headroom.
This page's figures change when the model or the runtime does.Last updated 2026-10-18 RSS → /feed.xml