Mac M3 Max 48G — what LLMs can it run?
88 of 110 indexed models fit comfortably in 48GB at 4K context, each at the highest-quality quant that still leaves headroom.
The short answer for Mac M3 Max 48G
48 GB, 400 GB/s Unified LPDDR5. 88 of 110 models in this index fit comfortably at 4K context; 90 load at all.
- Biggest that fits
- Nemotron 3.5 Bolt 74B-A9B — 23.1 GB at IQ2_XXS — 13.0 GB spare at 4K, so longer context comes out of a thin margin.
- Room to grow
- Nemotron 3.5 Lightning 30B-A3B — 19.8 GB at Q4_0 — under 60% of the card, which leaves 16.2 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. This index has 1 measured run on this card; see the table below for what they actually returned.
- The wall you will hit
- The GPU's share of unified memory, not the 48 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+ · 1
| Model | Quant | Est. VRAM | Headroom | tok/s |
|---|---|---|---|---|
| Nemotron 3.5 Bolt 74B-A9B74B-A9B | IQ2_XXS | 23.05 GB | +13 GB | ≤157Ceiling |
32B · 30
| Model | Quant | Est. VRAM | Headroom | tok/s |
|---|---|---|---|---|
| Kimi Linear 48B-A3B Instruct48B-A3B | Q4_K_M | 30.44 GB | +5.6 GB | ≤232Ceiling |
| Mixtral 8x7B Instruct47B MoE | Q4_K_M | 30.13 GB | +5.9 GB | ≤54Ceiling |
| Seed-OSS 36B Instruct36B | Q5_K_M | 27.81 GB | +8.2 GB | ≤16Ceiling |
| Qwen3.6 35B-A3B35B-A3B | Q4_K_M | 23.38 GB | +12.6 GB | ≤224Ceiling |
| Command R 35B35B | Q4_K_M | 22.86 GB | +13.1 GB | ≤20Ceiling |
| Yi 1.5 34B Chat34B | Q4_K_M | 22.82 GB | +13.2 GB | ≤20Ceiling |
| Qwen3 32B Instruct32B | Q4_K_M | 21.85 GB | +14.2 GB | 13–17Derived |
| Qwen2.5 32B Instruct32B | Q4_K_M | 21.69 GB | +14.3 GB | 13–17Derived |
| Qwen2.5-Coder 32B Instruct32B | Q4_K_M | 21.69 GB | +14.3 GB | 13–17Derived |
| DeepSeek-R1-Distill-Qwen-32B32B | Q4_K_M | 21.69 GB | +14.3 GB | 13–17Derived |
| Nemotron 3.5 Lightning 30B-A3B30B-A3B | Q4_0 | 19.79 GB | +16.2 GB | ≤234Ceiling |
| Gemma 4 31B IT31B | Q4_K_M | 21.08 GB | +14.9 GB | 13–18Derived |
| Qwen3 30B-A3B Instruct30B-A3B | Q5_K_M | 23.04 GB | +13 GB | ≤198Ceiling |
| Qwen3-Coder 30B-A3B Instruct30B-A3B | Q5_K_M | 23.04 GB | +13 GB | ≤198Ceiling |
| Qwen3-VL 30B-A3B Instruct30B-A3B | Q4_K_M | 19.73 GB | +16.3 GB | ≤231Ceiling |
| GLM-4.7-Flash30B-A3B | Q4_K_M | 19.19 GB | +16.8 GB | ≤232Ceiling |
| Granite 4.2 30B30B | Q6_K | 26.19 GB | +9.8 GB | 10–14Derived |
| Muse Glimmer 30B30B | Q8_0 | 31.08 GB | +4.9 GB | 8–11Derived |
| Qwen3.8 27B27B | Q4_K_M | 17.47 GB | +18.5 GB | 15–20Derived |
| Qwen3.6 27B27B | Q8_0 | 30.71 GB | +5.3 GB | 9–11Derived |
| Gemma 3 27B IT27B | Q4_K_M | 18.37 GB | +17.6 GB | 15–20Derived |
| Gemma 2 27B Instruct27B | Q5_K_M | 21.76 GB | +14.2 GB | 13–17Derived |
| Gemma 4 26B-A4B IT26B-A4B | Q8_0 | 28.42 GB | +7.6 GB | ≤99Ceiling |
| Mistral Small 24B Instruct24B | Q4_K_M | 15.89 GB | +20.1 GB | 17–23Derived |
| LFM2.5 24B-A2B24B-A2B | Q8_0 | 26.73 GB | +9.3 GB | ≤162Ceiling |
| Devstral Small 1.1 24B24B | Q6_K | 20.91 GB | +15.1 GB | 13–17Derived |
| Magistral Small 1.2 24B24B | Q6_K | 20.91 GB | +15.1 GB | 13–17Derived |
| Codestral 22B22B | Q4_K_M | 15.03 GB | +21 GB | 19–25Derived |
| ERNIE 4.5 21B-A3B21B-A3B | Q8_0 | 24.44 GB | +11.6 GB | ≤132Ceiling |
| GPT-OSS 20B21B MoE | MXFP4 | 12.76 GB | +23.2 GB | ≤202Ceiling |
14B · 18
| Model | Quant | Est. VRAM | Headroom | tok/s |
|---|---|---|---|---|
| InternLM2 20B Chat20B | Q5_K_M | 15.59 GB | +20.4 GB | 18–24Derived |
| DeepSeek-Coder-V2-Lite Instruct16B-A2.4B | Q8_0 | 17.56 GB | +18.4 GB | ≤165Ceiling |
| DeepSeek-V2-Lite Chat16B-A2.4B | Q4_K_M | 10.08 GB | +25.9 GB | ≤289Ceiling |
| StarCoder2 15B15B | Q4_K_M | 10.19 GB | +25.8 GB | 28–37Derived |
| Qwen3 14B Instruct14B | Q5_K_M | 11.65 GB | +24.4 GB | 24–32Derived |
| Qwen2.5 14B Instruct14B | Q5_K_M | 11.73 GB | +24.3 GB | 24–32Derived |
| DeepSeek-R1-Distill-Qwen-14B14B | Q4_K_M | 10.14 GB | +25.9 GB | 28–37Derived |
| Phi-4 14B14B | Q5_K_M | 11.77 GB | +24.2 GB | 24–32Derived |
| Phi-3 Medium 14B Instruct14B | Q6_K | 12.86 GB | +23.1 GB | 22–29Derived |
| Mistral Nemo 12B Instruct12B | Q6_K | 11.14 GB | +24.9 GB | 25–33Derived |
| Gemma 3 12B IT12B | Q5_K_M | 9.84 GB | +26.2 GB | 29–38Derived |
| Mellum2.1 12B-A2.5B Thinking12B-A2.5B | Q8_0 | 13.6 GB | +22.4 GB | ≤159Ceiling |
| Stable LM 2 12B Chat12B | Q4_K_M | 8.35 GB | +27.7 GB | 34–45Derived |
| Llama 3.2 11B Vision Instruct11B | Q8_0 | 12.9 GB | +23.1 GB | 21–28Derived |
| Solar 10.7B Instruct11B | Q4_K_M | 7.6 GB | +28.4 GB | 39–51Derived |
| Falcon 3 10B Instruct10B | Q4_K_M | 7.28 GB | +28.7 GB | 40–53Derived |
| Gemma 2 9B Instruct9B | Q8_0 | 11.7 GB | +24.3 GB | 26–34Derived |
| Qwen3.5 9B9B | Q8_0 | 10.37 GB | +25.6 GB | 26–34Derived |
7B · 25
| Model | Quant | Est. VRAM | Headroom | tok/s |
|---|---|---|---|---|
| GLM-4-9B-Chat9B | Q8_0 | 10.16 GB | +25.8 GB | 26–35Derived |
| Qwen3-VL 8B Instruct8B | Q8_0 | 10.39 GB | +25.6 GB | 27–36Derived |
| Granite 4.2 8B8B | Q8_0 | 10.45 GB | +25.6 GB | 27–36Derived |
| Ministral 3 8B Instruct8B | Q8_0 | 10.02 GB | +26 GB | 28–37Derived |
| Qwen2-VL 7B Instruct7B | Q4_K_M | 5.49 GB | +30.5 GB | 50–66Derived |
| Granite 3.1 8B Instruct8B | Q4_K_M | 5.88 GB | +30.1 GB | 51–67Derived |
| Qwen3 8B Instruct8B | Q6_K | 7.64 GB | +28.4 GB | 37–49Derived |
| Llama 3.1 8B Instruct8B | Q8_0 | 9.47 GB | +26.5 GB | 29–39Derived |
| Nous Hermes 3 Llama 3.1 8B8B | Q4_K_M | 5.64 GB | +30.4 GB | 52–68Derived |
| Aya 23 8B8B | Q4_K_M | 5.64 GB | +30.4 GB | 52–68Derived |
| OpenChat 3.6 8B8B | Q4_K_M | 5.64 GB | +30.4 GB | 52–68Derived |
| DeepSeek-R1-Distill-Llama-8B8B | Q5_K_M | 6.51 GB | +29.5 GB | 44–58Derived |
| InternLM2 7B Chat7B | Q4_K_M | 5.45 GB | +30.6 GB | 53–71Derived |
| Qwen2.5 7B Instruct7B | Q6_K | 6.77 GB | +29.2 GB | 40–53Derived |
| Qwen2.5-Coder 7B Instruct7B | Q4_K_M | 5.07 GB | +30.9 GB | 54–72Derived |
| DeepSeek-R1-Distill-Qwen-7B7B | Q4_K_M | 5.07 GB | +30.9 GB | 54–72Derived |
| Gemma 4 E4B ITE4B | Q8_0 | 8.46 GB | +27.5 GB | 31–42Derived |
| OLMo 2 7B Instruct7B | Q8_0 | 10.3 GB | +25.7 GB | 32–43Derived |
| OLMo 3 7B Instruct7B | Q8_0 | 10.3 GB | +25.7 GB | 32–43Derived |
| WizardLM-2 7B7B | Q4_K_M | 5.14 GB | +30.9 GB | 57–76Derived |
| Mistral 7B Instruct v0.37B | Q6_K | 6.75 GB | +29.3 GB | 42–56Derived |
| Zephyr 7B Beta7B | Q6_K | 6.75 GB | +29.3 GB | 42–56Derived |
| Gemma 4 E2B ITE2B | Q8_0 | 5.2 GB | +30.8 GB | ≤85Ceiling |
| Qwen3.5 4B4B | Q8_0 | 4.97 GB | +31 GB | ≤91Ceiling |
| Gemma 3 4B IT4B | Q8_0 | 5.05 GB | +31 GB | ≤92Ceiling |
≤3B · 14
| Model | Quant | Est. VRAM | Headroom | tok/s |
|---|---|---|---|---|
| Qwen3 4B Instruct4B | Q6_K | 4.05 GB | +32 GB | ≤128Ceiling |
| Phi-4 Mini Instruct3.8B | Q8_0 | 4.81 GB | +31.2 GB | ≤103Ceiling |
| Phi-3.5 Mini Instruct3.8B | Q8_0 | 5.89 GB | +30.1 GB | ≤104Ceiling |
| Granite 4.2 3B3B | Q8_0 | 4.41 GB | +31.6 GB | ≤108Ceiling |
| Llama 3.2 3B Instruct3B | Q8_0 | 4.05 GB | +32 GB | ≤123Ceiling |
| Qwen2.5 3B Instruct3B | Q8_0 | 3.59 GB | +32.4 GB | ≤128Ceiling |
| SmolLM3 3B3B | Q8_0 | 3.73 GB | +32.3 GB | ≤129Ceiling |
| Gemma 2 2B Instruct2B | Q8_0 | 3.34 GB | +32.7 GB | ≤152Ceiling |
| Qwen3.5 2B2B | Q8_0 | 2.23 GB | +33.8 GB | ≤202Ceiling |
| Qwen3 1.7B Instruct1.7B | Q8_0 | 2.39 GB | +33.6 GB | ≤230Ceiling |
| Qwen2.5 1.5B Instruct1.5B | Q8_0 | 1.83 GB | +34.2 GB | ≤258Ceiling |
| Llama 3.2 1B Instruct1B | Q8_0 | 1.51 GB | +34.5 GB | ≤320Ceiling |
| Qwen3.5 0.8B0.8B | Q8_0 | 0.94 GB | +35.1 GB | ≤494Ceiling |
| Qwen2.5 0.5B Instruct0.5B | Q8_0 | 0.6 GB | +35.4 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
1 benchmark runs in this index were recorded on a Mac M3 Max 48G. Everything else on this page is calculated, not observed.
| Model | Stack | tok/s | VRAM used |
|---|---|---|---|
| Llama 3.1 8B | Ollama · GGUF Q4_K_M | 68 | 5.7 GB |
Cards with the same budget
What fits is decided by memory, so every 48GB card of this type returns the same list. These pages are not different answers — they differ in throughput, which this index does not measure per card.
Same list, different speed — play one sentence on both cards: Mac M3 Max 48G vs Mac M4 Pro 48G (Qwen3 32B Instruct)
Setup guides for this card
Step-by-step guides written for the Mac M3 Max 48G'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 Pro 64G (64GB) fits 2 more of the indexed models than this card. Mac M5 Pro 64G →
Common questions
What is the best local LLM for a Mac M3 Max 48G?
For everyday use, Nemotron 3.5 Lightning 30B-A3B at Q4_0 — about 19.8 GB of the card's 48 GB at 4K context, leaving 16.2 GB for a longer window. If you want the largest thing that will load, that is Nemotron 3.5 Bolt 74B-A9B at IQ2_XXS (23.1 GB, 13.0 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 M3 Max 48G run Llama 3.2 90B Vision Instruct?
No. Its smallest build here, Q3_K_M, needs about 46.2 GB and this card has 36.0 GB usable — short by 10.2 GB before any context beyond 4K. The largest model this card does clear is Nemotron 3.5 Bolt 74B-A9B.
How many tokens per second does a Mac M3 Max 48G do on an 8B model at Q4?
This index has 1 measured run on this card — those figures are in the table on this page. 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 M3 Max 48G or L40S 48G for local LLMs?
They do not hold the same set: Mac M3 Max 48G fits 88 and L40S 48G 94 of 110 at 4K — macOS keeps part of the Mac M3 Max 48G's unified memory from the GPU, leaving about 36 GB usable. Speed is the other difference — 400 GB/s against 864 GB/s, a 2.2× gap in how fast the weights can be read, which is what token generation is bound by. The L40S 48G generates faster on any model both can hold.
88 of 110 indexed models fit comfortably in 48GB 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