Mac M3 8G — what LLMs can it run?
25 of 110 indexed models fit comfortably in 8GB at 4K context, each at the highest-quality quant that still leaves headroom.
The short answer for Mac M3 8G
8 GB, 100 GB/s Unified LPDDR5. 25 of 110 models in this index fit comfortably at 4K context; 37 load at all.
- Biggest that fits
- Llama 3.1 8B Instruct — 3.3 GB at Q2_K — 2.7 GB spare at 4K, so longer context comes out of a thin margin.
- 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 8 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.
7B · 11
| Model | Quant | Est. VRAM | Headroom | tok/s |
|---|---|---|---|---|
| Llama 3.1 8B Instruct8B | Q2_K | 3.31 GB | +2.7 GB | 31–40Derived |
| Qwen2.5 7B Instruct7B | Q4_K_M | 5.07 GB | +0.9 GB | 17–23Derived |
| Qwen2.5-Coder 7B Instruct7B | Q4_K_M | 5.07 GB | +0.9 GB | 17–23Derived |
| DeepSeek-R1-Distill-Qwen-7B7B | Q4_K_M | 5.07 GB | +0.9 GB | 17–23Derived |
| Gemma 4 E4B ITE4B | Q4_K_M | 4.88 GB | +1.1 GB | 18–23Derived |
| WizardLM-2 7B7B | Q4_K_M | 5.14 GB | +0.9 GB | 18–24Derived |
| Mistral 7B Instruct v0.37B | Q4_K_M | 5.14 GB | +0.9 GB | 18–24Derived |
| Zephyr 7B Beta7B | Q4_K_M | 5.14 GB | +0.9 GB | 18–24Derived |
| Gemma 4 E2B ITE2B | Q8_0 | 5.2 GB | +0.8 GB | ≤21Ceiling |
| Qwen3.5 4B4B | Q8_0 | 4.97 GB | +1 GB | ≤23Ceiling |
| Gemma 3 4B IT4B | Q8_0 | 5.05 GB | +1 GB | ≤23Ceiling |
≤3B · 14
| Model | Quant | Est. VRAM | Headroom | tok/s |
|---|---|---|---|---|
| Qwen3 4B Instruct4B | Q6_K | 4.05 GB | +2 GB | ≤32Ceiling |
| Phi-4 Mini Instruct3.8B | Q8_0 | 4.81 GB | +1.2 GB | ≤26Ceiling |
| Phi-3.5 Mini Instruct3.8B | Q4_K_M | 4.07 GB | +1.9 GB | ≤45Ceiling |
| Granite 4.2 3B3B | Q8_0 | 4.41 GB | +1.6 GB | ≤27Ceiling |
| Llama 3.2 3B Instruct3B | Q8_0 | 4.05 GB | +2 GB | ≤31Ceiling |
| Qwen2.5 3B Instruct3B | Q8_0 | 3.59 GB | +2.4 GB | ≤32Ceiling |
| SmolLM3 3B3B | Q8_0 | 3.73 GB | +2.3 GB | ≤32Ceiling |
| Gemma 2 2B Instruct2B | Q8_0 | 3.34 GB | +2.7 GB | ≤38Ceiling |
| Qwen3.5 2B2B | Q8_0 | 2.23 GB | +3.8 GB | ≤51Ceiling |
| Qwen3 1.7B Instruct1.7B | Q8_0 | 2.39 GB | +3.6 GB | ≤57Ceiling |
| Qwen2.5 1.5B Instruct1.5B | Q8_0 | 1.83 GB | +4.2 GB | ≤65Ceiling |
| Llama 3.2 1B Instruct1B | Q8_0 | 1.51 GB | +4.5 GB | ≤80Ceiling |
| Qwen3.5 0.8B0.8B | Q8_0 | 0.94 GB | +5.1 GB | ≤123Ceiling |
| Qwen2.5 0.5B Instruct0.5B | Q8_0 | 0.6 GB | +5.4 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 12 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 8G. 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 8GB 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 8G vs Mac M1 8G (Llama 3.1 8B Instruct)
Setup guides for this card
Step-by-step guides written for the Mac M3 8G'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 16G (16GB) fits 31 more of the indexed models than this card. Mac M3 16G →
Before buying, hear the speed difference on a model both cards hold: Mac M3 8G vs Mac M3 16G (Llama 3.1 8B Instruct) →
Common questions
What is the best local LLM for a Mac M3 8G?
For everyday use, Llama 3.1 8B Instruct at Q2_K — about 3.3 GB of the card's 8 GB at 4K context, leaving 2.7 GB for a longer window. If you want the largest thing that will load, that is Llama 3.1 8B Instruct at Q2_K (3.3 GB, 2.7 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 8G run Qwen3 8B Instruct?
Only just. Its smallest build here, Q4_K_M, needs about 5.8 GB against 6.0 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 M3 8G 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 8G or RTX 3070 Ti for local LLMs?
They do not hold the same set: Mac M3 8G fits 25 and RTX 3070 Ti 44 of 110 at 4K — macOS keeps part of the Mac M3 8G's unified memory from the GPU, leaving about 6 GB usable. Speed is the other difference — 100 GB/s against 608 GB/s, a 6.1× gap in how fast the weights can be read, which is what token generation is bound by. The RTX 3070 Ti generates faster on any model both can hold.
25 of 110 indexed models fit comfortably in 8GB 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