Mac M3 16G — what LLMs can it run?
56 of 110 indexed models fit comfortably in 16GB at 4K context, each at the highest-quality quant that still leaves headroom.
The short answer for Mac M3 16G
16 GB, 100 GB/s Unified LPDDR5. 56 of 110 models in this index fit comfortably at 4K context; 57 load at all.
- Biggest that fits
- DeepSeek-Coder-V2-Lite Instruct — 10.1 GB at Q4_K_M — 1.9 GB spare at 4K, so longer context comes out of a thin margin.
- Room to grow
- Qwen2-VL 7B Instruct — 5.5 GB at Q4_K_M — under 60% of the card, which leaves 6.5 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 100 GB/s puts a hard ceiling near 22 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 16 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 100 GB/s, which is the number that decides tok/s once a model fits.
14B · 17
| Model | Quant | Est. VRAM | Headroom | tok/s |
|---|---|---|---|---|
| DeepSeek-Coder-V2-Lite Instruct16B-A2.4B | Q4_K_M | 10.08 GB | +1.9 GB | ≤72Ceiling |
| DeepSeek-V2-Lite Chat16B-A2.4B | Q4_K_M | 10.08 GB | +1.9 GB | ≤72Ceiling |
| StarCoder2 15B15B | Q4_K_M | 10.19 GB | +1.8 GB | 9–12Derived |
| Qwen3 14B Instruct14B | Q4_K_M | 10.04 GB | +2 GB | 9–12Derived |
| Qwen2.5 14B Instruct14B | Q4_K_M | 10.14 GB | +1.9 GB | 9–12Derived |
| DeepSeek-R1-Distill-Qwen-14B14B | Q4_K_M | 10.14 GB | +1.9 GB | 9–12Derived |
| Phi-4 14B14B | Q4_K_M | 10.17 GB | +1.8 GB | 9–12Derived |
| Phi-3 Medium 14B Instruct14B | Q4_K_M | 9.73 GB | +2.3 GB | 10–13Derived |
| Mistral Nemo 12B Instruct12B | Q4_K_M | 8.42 GB | +3.6 GB | 11–14Derived |
| Gemma 3 12B IT12B | Q5_K_M | 9.84 GB | +2.2 GB | 9–12Derived |
| Mellum2.1 12B-A2.5B Thinking12B-A2.5B | Q4_K_M | 8.59 GB | +3.4 GB | ≤63Ceiling |
| Stable LM 2 12B Chat12B | Q4_K_M | 8.35 GB | +3.7 GB | 11–15Derived |
| Llama 3.2 11B Vision Instruct11B | Q4_K_M | 7.66 GB | +4.3 GB | 12–16Derived |
| Solar 10.7B Instruct11B | Q4_K_M | 7.6 GB | +4.4 GB | 12–16Derived |
| Falcon 3 10B Instruct10B | Q4_K_M | 7.28 GB | +4.7 GB | 13–17Derived |
| Gemma 2 9B Instruct9B | Q4_K_M | 7.3 GB | +4.7 GB | 14–19Derived |
| Qwen3.5 9B9B | Q8_0 | 10.37 GB | +1.6 GB | 8–11Derived |
7B · 25
| Model | Quant | Est. VRAM | Headroom | tok/s |
|---|---|---|---|---|
| GLM-4-9B-Chat9B | Q8_0 | 10.16 GB | +1.8 GB | 8–11Derived |
| Qwen3-VL 8B Instruct8B | Q8_0 | 10.39 GB | +1.6 GB | 9–11Derived |
| Granite 4.2 8B8B | Q8_0 | 10.45 GB | +1.6 GB | 9–11Derived |
| Ministral 3 8B Instruct8B | Q8_0 | 10.02 GB | +2 GB | 9–12Derived |
| Qwen2-VL 7B Instruct7B | Q4_K_M | 5.49 GB | +6.5 GB | 16–21Derived |
| Granite 3.1 8B Instruct8B | Q4_K_M | 5.88 GB | +6.1 GB | 16–22Derived |
| Qwen3 8B Instruct8B | Q6_K | 7.64 GB | +4.4 GB | 12–16Derived |
| Llama 3.1 8B Instruct8B | Q8_0 | 9.47 GB | +2.5 GB | 9–13Derived |
| Nous Hermes 3 Llama 3.1 8B8B | Q4_K_M | 5.64 GB | +6.4 GB | 17–22Derived |
| Aya 23 8B8B | Q4_K_M | 5.64 GB | +6.4 GB | 17–22Derived |
| OpenChat 3.6 8B8B | Q4_K_M | 5.64 GB | +6.4 GB | 17–22Derived |
| DeepSeek-R1-Distill-Llama-8B8B | Q5_K_M | 6.51 GB | +5.5 GB | 14–19Derived |
| InternLM2 7B Chat7B | Q4_K_M | 5.45 GB | +6.6 GB | 17–23Derived |
| Qwen2.5 7B Instruct7B | Q6_K | 6.77 GB | +5.2 GB | 13–17Derived |
| Qwen2.5-Coder 7B Instruct7B | Q4_K_M | 5.07 GB | +6.9 GB | 17–23Derived |
| DeepSeek-R1-Distill-Qwen-7B7B | Q4_K_M | 5.07 GB | +6.9 GB | 17–23Derived |
| Gemma 4 E4B ITE4B | Q8_0 | 8.46 GB | +3.5 GB | 10–13Derived |
| OLMo 2 7B Instruct7B | Q8_0 | 10.3 GB | +1.7 GB | 10–14Derived |
| OLMo 3 7B Instruct7B | Q8_0 | 10.3 GB | +1.7 GB | 10–14Derived |
| WizardLM-2 7B7B | Q4_K_M | 5.14 GB | +6.9 GB | 18–24Derived |
| Mistral 7B Instruct v0.37B | Q6_K | 6.75 GB | +5.3 GB | 14–18Derived |
| Zephyr 7B Beta7B | Q6_K | 6.75 GB | +5.3 GB | 14–18Derived |
| Gemma 4 E2B ITE2B | Q8_0 | 5.2 GB | +6.8 GB | ≤21Ceiling |
| Qwen3.5 4B4B | Q8_0 | 4.97 GB | +7 GB | ≤23Ceiling |
| Gemma 3 4B IT4B | Q8_0 | 5.05 GB | +7 GB | ≤23Ceiling |
≤3B · 14
| Model | Quant | Est. VRAM | Headroom | tok/s |
|---|---|---|---|---|
| Qwen3 4B Instruct4B | Q6_K | 4.05 GB | +8 GB | ≤32Ceiling |
| Phi-4 Mini Instruct3.8B | Q8_0 | 4.81 GB | +7.2 GB | ≤26Ceiling |
| Phi-3.5 Mini Instruct3.8B | Q8_0 | 5.89 GB | +6.1 GB | ≤26Ceiling |
| Granite 4.2 3B3B | Q8_0 | 4.41 GB | +7.6 GB | ≤27Ceiling |
| Llama 3.2 3B Instruct3B | Q8_0 | 4.05 GB | +8 GB | ≤31Ceiling |
| Qwen2.5 3B Instruct3B | Q8_0 | 3.59 GB | +8.4 GB | ≤32Ceiling |
| SmolLM3 3B3B | Q8_0 | 3.73 GB | +8.3 GB | ≤32Ceiling |
| Gemma 2 2B Instruct2B | Q8_0 | 3.34 GB | +8.7 GB | ≤38Ceiling |
| Qwen3.5 2B2B | Q8_0 | 2.23 GB | +9.8 GB | ≤51Ceiling |
| Qwen3 1.7B Instruct1.7B | Q8_0 | 2.39 GB | +9.6 GB | ≤57Ceiling |
| Qwen2.5 1.5B Instruct1.5B | Q8_0 | 1.83 GB | +10.2 GB | ≤65Ceiling |
| Llama 3.2 1B Instruct1B | Q8_0 | 1.51 GB | +10.5 GB | ≤80Ceiling |
| Qwen3.5 0.8B0.8B | Q8_0 | 0.94 GB | +11.1 GB | ≤123Ceiling |
| Qwen2.5 0.5B Instruct0.5B | Q8_0 | 0.6 GB | +11.4 GB | ≤204Ceiling |
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 1 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 M3 16G. Every figure on this page is calculated from the model architecture and the quant level — treat them as estimates, not measurements.
Cards with the same budget
What fits is decided by memory, so every 16GB 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.
The same 16GB in system memory is a different machine — everything on this page fits there, at a speed set by your DIMMs rather than a graphics bus: 16 GB RAM (CPU)
Same list, different speed — play one sentence on both cards: Mac M3 16G vs Mac M4 16G (Qwen3 14B Instruct) · Mac M3 16G vs Mac M2 16G (Qwen3 14B Instruct) · Mac M3 16G vs Mac M1 16G (Qwen3 14B Instruct)
Setup guides for this card
Step-by-step guides written for the Mac M3 16G's runtime or memory size, with the commands and the checks that prove the model is running on the card.
Stepping up
A Mac M3 Pro 18G (18GB) fits 1 more of the indexed models than this card. Mac M3 Pro 18G →
Before buying, hear the speed difference on a model both cards hold: Mac M3 16G vs Mac M3 Pro 18G (Qwen3 14B Instruct) →
Common questions
What is the best local LLM for a Mac M3 16G?
For everyday use, Qwen2-VL 7B Instruct at Q4_K_M — about 5.5 GB of the card's 16 GB at 4K context, leaving 6.5 GB for a longer window. If you want the largest thing that will load, that is DeepSeek-Coder-V2-Lite Instruct at Q4_K_M (10.1 GB, 1.9 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 16G run InternLM2 20B Chat?
No. Its smallest build here, Q4_K_M, needs about 13.4 GB and this card has 12.0 GB usable — short by 1.4 GB before any context beyond 4K. The largest model this card does clear is DeepSeek-Coder-V2-Lite Instruct.
How many tokens per second does a Mac M3 16G 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 100 GB/s — a ceiling near 22 tok/s. Batch size, context length, the runtime and how much of the model sits in cache all take you below it.
Mac M3 16G or RTX 5080 for local LLMs?
They do not hold the same set: Mac M3 16G fits 56 and RTX 5080 63 of 110 at 4K — macOS keeps part of the Mac M3 16G's unified memory from the GPU, leaving about 12 GB usable. Speed is the other difference — 100 GB/s against 960 GB/s, a 9.6× gap in how fast the weights can be read, which is what token generation is bound by. The RTX 5080 generates faster on any model both can hold.
56 of 110 indexed models fit comfortably in 16GB 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