Which local LLMs can the Apple M5 Max (32-core GPU) run?
Unified memory — by default the GPU can use about 70% of it
Assumes 8k context and an f16 KV cache. Verdicts are shown for each memory size.
Specs
- Memory options
- 36 GB · 48 GB
- Memory bandwidth
- 460.0 GB/s
- FP16 compute
- 30.0 TFLOPS
- Launch year
- 2026
Verdicts at a glance
At Q4_K_M (or the closest available quant) with 8k context.
- 36 GBRuns great15 modelsRuns well9 modelsRuns slowly0 modelsWon't run5 models1 more model runs at a lower quant.
- 48 GBRuns great15 modelsRuns well9 modelsRuns slowly0 modelsWon't run5 models3 more models run at a lower quant.
How to read the verdicts
- Runs great
- Fully on the GPU at 20 tok/s or more
- Runs well
- Fully on the GPU at 8–20 tok/s; MoE experts in system RAM at 20 tok/s or more; or at least 90% on the GPU at 8 tok/s or more
- Runs slowly
- 2–8 tok/s; CPU-only; less than 90% on the GPU; or MoE experts in system RAM below 20 tok/s
- Won't run
- Does not fit, or under 2 tok/s
Every model on the Apple M5 Max (32-core GPU)
Scroll sideways to see every column.
| Model | Verdict | Speed | Quant | Memory | Runs as | Context | Notes |
|---|---|---|---|---|---|---|---|
| EXAONE 4.0 1.2B | Runs great | est. 210.2 tok/s168.1–252.2calibrated estimate ±20% | Q4_K_M | 1.9 / 33.6 GB | Unified memory | up to 64k | — |
| HyperCLOVA X SEED 1.5B | Runs great | est. 152.0 tok/s121.6–182.4calibrated estimate ±20% | Q4_K_M | 2.4 / 33.6 GB | Unified memory | up to 16k | — |
| Qwen3.5 35B-A3B | Runs great | est. 67.1 tok/s53.7–80.5calibrated estimate ±20% | Q4_K_M | 23.4 / 33.6 GB | Unified memory | up to 128k | — |
| gpt-oss-20b | Runs great | est. 60.3 tok/s48.3–72.4calibrated estimate ±20% | MXFP4 | 12.9 / 33.6 GB | Unified memory | up to 64k |
|
| Qwen3 30B-A3B (2507) | Runs great | est. 49.1 tok/s39.2–58.9calibrated estimate ±20% | Q4_K_M | 19.9 / 33.6 GB | Unified memory | up to 32k | — |
| Gemma 4 26B-A4B | Runs great | est. 47.7 tok/s38.2–57.3calibrated estimate ±20% | Q4_K_M | 17.9 / 33.6 GB | Unified memory | up to 64k | — |
| DeepSeek R1 Distill Llama 8B | Runs great | est. 46.0 tok/s36.8–55.3calibrated estimate ±20% | Q4_K_M | 6.6 / 33.6 GB | Unified memory | up to 16k |
|
| Kanana 1.5 8B | Runs great | est. 46.0 tok/s36.8–55.3calibrated estimate ±20% | Q4_K_M | 6.6 / 33.6 GB | Unified memory | up to 32k | — |
| Llama 3.1 8B | Runs great | est. 46.0 tok/s36.8–55.3calibrated estimate ±20% | Q4_K_M | 6.6 / 33.6 GB | Unified memory | up to 64k | — |
| Kanana 1.5 15.7B-A3B | Runs great | est. 45.0 tok/s36.0–54.0calibrated estimate ±20% | Q4_K_M | 12.1 / 33.6 GB | Unified memory | up to 32k | — |
| Qwen3.5 9B | Runs great | est. 45.0 tok/s36.0–54.0calibrated estimate ±20% | Q4_K_M | 6.7 / 33.6 GB | Unified memory | up to 128k | — |
| Qwen3 8B | Runs great | est. 44.2 tok/s35.4–53.1calibrated estimate ±20% | Q4_K_M | 6.8 / 33.6 GB | Unified memory | up to 32k | — |
| Gemma 4 12B | Runs great | est. 36.0 tok/s28.8–43.3calibrated estimate ±20% | Q4_K_M | 8.2 / 33.6 GB | Unified memory | up to 64k | — |
| Gemma 3 12B | Runs great | est. 35.2 tok/s28.2–42.3calibrated estimate ±20% | Q4_K_M | 8.4 / 33.6 GB | Unified memory | up to 64k | — |
| Qwen3 14B | Runs great | est. 26.7 tok/s21.3–32.0calibrated estimate ±20% | Q4_K_M | 10.9 / 33.6 GB | Unified memory | up to 16k | — |
| HyperCLOVA X SEED Think 14B | Runs well | est. 27.1 tok/s19.0–35.2theoretical estimate ±30% | Q4_K_M | 10.8 / 33.6 GB | Unified memory | up to 64k |
|
| Phi-4 | Runs well | est. 25.7 tok/s20.6–30.9calibrated estimate ±20% | Q4_K_M | 11.3 / 33.6 GB | Unified memory | up to 16k |
|
| Mistral Small 3.2 24B | Runs well | est. 17.6 tok/s14.1–21.1calibrated estimate ±20% | Q4_K_M | 16.2 / 33.6 GB | Unified memory | up to 64k |
|
| Gemma 3 27B | Runs well | est. 16.0 tok/s12.8–19.2calibrated estimate ±20% | Q4_K_M | 17.8 / 33.6 GB | Unified memory | up to 128k |
|
| Qwen3.5 27B | Runs well | est. 15.6 tok/s12.5–18.8calibrated estimate ±20% | Q4_K_M | 18.2 / 33.6 GB | Unified memory | up to 128k |
|
| EXAONE 4.0 32B | Runs well | est. 13.9 tok/s11.1–16.7calibrated estimate ±20% | Q4_K_M | 20.5 / 33.6 GB | Unified memory | up to 32k |
|
| EXAONE 4.5 33B | Runs well | est. 13.4 tok/s10.7–16.1calibrated estimate ±20% | Q4_K_M | 21.2 / 33.6 GB | Unified memory | up to 32k |
|
| Qwen3 32B | Runs well | est. 12.6 tok/s10.1–15.1calibrated estimate ±20% | Q4_K_M | 22.5 / 33.6 GB | Unified memory | up to 32k | — |
| DeepSeek R1 Distill Qwen 32B | Runs well | est. 12.5 tok/s10.0–15.1calibrated estimate ±20% | Q4_K_M | 22.6 / 33.6 GB | Unified memory | up to 8k |
|
| Qwen3.5 122B-A10B | Won't runTry UD-Q2_K_XL (heavy quality loss): Runs slowly | est. 38.0 tok/s26.6–49.5theoretical estimate ±30% | UD-Q2_K_XL | 43.1 GB RAM | CPU only | up to 32k |
|
| Solar Open 100B | Won't runTry Q2_K (heavy quality loss): Runs slowly | est. 22.9 tok/s16.1–29.8theoretical estimate ±30% | Q2_K | 39.3 GB RAM | CPU only | up to 16k |
|
| Llama 3.3 70B | Won't runTry IQ4_XS: Runs slowly | est. 4.0 tok/s2.8–5.1theoretical estimate ±30% | IQ4_XS | 40.6 GB RAM | CPU only | up to 16k |
|
| gpt-oss-120b | Won't run | — | MXFP4 | needs 64.3 GB | — | — |
|
| Solar Open 2 250B | Won't run | — | IQ4_XS | needs 137.2 GB | — | — |
|
Reasoning models spend extra tokens thinking, so their speed thresholds are 1.5× stricter (30 / 12 / 3 tok/s).
Verdict by memory size
| Model | 36 GB | 48 GB |
|---|---|---|
| EXAONE 4.0 1.2B | Runs great | Runs great |
| HyperCLOVA X SEED 1.5B | Runs great | Runs great |
| Qwen3.5 35B-A3B | Runs great | Runs great |
| gpt-oss-20b | Runs great | Runs great |
| Qwen3 30B-A3B (2507) | Runs great | Runs great |
| Gemma 4 26B-A4B | Runs great | Runs great |
| DeepSeek R1 Distill Llama 8B | Runs great | Runs great |
| Kanana 1.5 8B | Runs great | Runs great |
| Llama 3.1 8B | Runs great | Runs great |
| Kanana 1.5 15.7B-A3B | Runs great | Runs great |
| Qwen3.5 9B | Runs great | Runs great |
| Qwen3 8B | Runs great | Runs great |
| Gemma 4 12B | Runs great | Runs great |
| Gemma 3 12B | Runs great | Runs great |
| Qwen3 14B | Runs great | Runs great |
| HyperCLOVA X SEED Think 14B | Runs well | Runs well |
| Phi-4 | Runs well | Runs well |
| Mistral Small 3.2 24B | Runs well | Runs well |
| Gemma 3 27B | Runs well | Runs well |
| Qwen3.5 27B | Runs well | Runs well |
| EXAONE 4.0 32B | Runs well | Runs well |
| EXAONE 4.5 33B | Runs well | Runs well |
| Qwen3 32B | Runs well | Runs well |
| DeepSeek R1 Distill Qwen 32B | Runs well | Runs well |
| Qwen3.5 122B-A10B | Won't run | Won't run |
| Solar Open 100B | Won't run | Won't run |
| Llama 3.3 70B | Won't run | Won't run |
| gpt-oss-120b | Won't run | Won't run |
| Solar Open 2 250B | Won't run | Won't run |
Measured results on the Apple M5 Max (32-core GPU)
No public measurements for this device yet.
Frequently asked questions
What is the largest model that runs entirely on the Apple M5 Max (32-core GPU)?
Qwen3.5 35B-A3B at Q4_K_M (a 22.63 GB file) fits entirely in 48 GB of unified memory with 8k context, at est. 67.1 tok/s (53.7–80.5, calibrated estimate ±20%).
How many local LLMs run well on the Apple M5 Max (32-core GPU)?
At Q4_K_M with 8k context, with 48 GB of memory, out of 29 tracked models: 15 run great, 9 run well, 0 run slowly and 5 won't run.
Can the Apple M5 Max (32-core GPU) run a 70B model like Llama 3.3 70B?
Runs slowly — CPU only, IQ4_XS: est. 4.0 tok/s (2.8–5.1, theoretical estimate ±30%).
Speeds are estimates from memory bandwidth, calibrated against public benchmarks, and each one comes with an error band and a confidence label. Real results vary with drivers, backend, context length and thermals.