Apple M3 Max (30-core GPU) for local LLMs
Written by Jakub Rusinowski · Last updated
With 96 GB at 300 GB/s, the M3 Max (30-core GPU) runs 136 catalogued models at Q4_K_M with 8K context. The largest that fits is GPT-OSS 120B (~71.6 GB), and the top pick is Qwen3-Coder-Next (80B-A3B MoE) at 41–86 tok/s.
The 14-core CPU / 30-core GPU M3 Max. Fewer memory controllers than the 40-core part: 300 GB/s rather than 400, and only 36 GB or 96 GB configurations. Modelling a 36 GB machine on the 40-core record overstates its throughput by a third.
Models that run on the M3 Max (30-core GPU)
Q4_K_M, 8K context, 72 GB usable. Ranked by quality and speed.
| Model | Memory · marker = 72 GB | VRAM | Speed |
|---|---|---|---|
| Qwen3-Coder-Next (80B-A3B MoE) Qwen3-Coder | 49.4 GB | 41–86 tok/s | |
| Kolibri 1 Kolibri | 48.5 GB | 37–78 tok/s | |
| Nemotron 3.5 Bolt 74B Nemotron 3.5 | 45.6 GB | 18–37 tok/s | |
| K2 Horizon MoVA 36B-A4B K2 Horizon | 25 GB | 24–49 tok/s | |
| Qwen 3.6 35B-A3B Qwen 3.6 | 22.2 GB | 42–87 tok/s | |
| Nex-N2.5 mini Nex-N2.5 | 21.9 GB | 26–53 tok/s | |
| Nex-N2 mini Nex-N2 | 21.9 GB | 26–53 tok/s | |
| Qwen 3.5 35B-A3B Qwen 3.5 | 21.9 GB | 26–53 tok/s | |
| Laguna XS 2.1 33B-A3B Poolside Laguna XS 2.1 | 20.7 GB | 26–53 tok/s | |
| Granite 4.0 Small-H 32B-A9B IBM Granite 4.0 | 20.1 GB | 14–29 tok/s | |
| Nemotron-Cascade 2 30B-A3B Nemotron Cascade 2 | 19.9 GB | 24–50 tok/s | |
| Qwen 3 30B-A3B (MoE) Qwen 3 | 20 GB | 32–67 tok/s |
Buy it or rent the same memory
Buy the card, or rent a GPU with the same memory by the hour to try models first.
As an Amazon Associate we earn from qualifying purchases. Cloud GPU links are referral links — we may earn a commission at no extra cost to you.
Speed vs other GPUs
Llama 3.1 8B, Q4_K_M. Estimated ranges. How this is calculated
Specifications
Specs last updated 2026-09-19.
- Memory
- 96 GB
- Memory bandwidth
- 300 GB/s
- Architecture
- ARM, 3nm TSMC
- Series
- Apple Silicon
- Board power
- 35 W
- Release year
- 2023
- Compute backends
- METAL
- Usable for models
- 72 GB (~75% of unified)
Similar GPUs
Frequently asked questions
Can the Apple M3 Max (30-core GPU) run local LLMs?
Yes. With 96 GB (72 GB usable by a model) it runs 136 of the catalogued models at Q4_K_M with 8K context; the largest is GPT-OSS 120B, needing about 71.6 GB.
How fast is the Apple M3 Max (30-core GPU) for AI inference?
It is estimated to run Llama 3.1 8B at 18–38 tok/s at Q4_K_M. Llama 3.3 70B is estimated at 2.4–5.0 tok/s. These are modelled estimates from memory bandwidth, not measurements; the methodology page shows the formula.
What LLMs can I run on 96 GB?
Among the best that fit: Qwen3-Coder-Next (80B-A3B MoE), Kolibri 1, Nemotron 3.5 Bolt 74B, K2 Horizon MoVA 36B-A4B, Qwen 3.6 35B-A3B.