Which local LLMs can the Apple M5 Max (40-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
- 48 GB · 64 GB · 128 GB
- Memory bandwidth
- 614.0 GB/s
- FP16 compute
- 38.0 TFLOPS
- Launch year
- 2026
Verdicts at a glance
At Q4_K_M (or the closest available quant) with 8k context.
- 48 GBRuns great20 modelsRuns well4 modelsRuns slowly0 modelsWon't run5 models3 more models run at a lower quant.
- 64 GBRuns great20 modelsRuns well4 modelsRuns slowly1 modelWon't run4 models2 more models run at a lower quant.
- 128 GBRuns great23 modelsRuns well5 modelsRuns slowly0 modelsWon't run1 model1 more model runs 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 (40-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. 280.5 tok/s224.4–336.6calibrated estimate ±20% | Q4_K_M | 1.9 / 89.6 GB | Unified memory | up to 64k | — |
| HyperCLOVA X SEED 1.5B | Runs great | est. 202.9 tok/s162.4–243.5calibrated estimate ±20% | Q4_K_M | 2.4 / 89.6 GB | Unified memory | up to 16k | — |
| Qwen3.5 35B-A3B | Runs great | est. 89.6 tok/s71.7–107.5calibrated estimate ±20% | Q4_K_M | 23.4 / 89.6 GB | Unified memory | up to 256k | — |
| gpt-oss-20b | Runs great | est. 80.5 tok/s64.4–96.6calibrated estimate ±20% | MXFP4 | 12.9 / 89.6 GB | Unified memory | up to 128k |
|
| Qwen3 30B-A3B (2507) | Runs great | est. 65.5 tok/s52.4–78.6calibrated estimate ±20% | Q4_K_M | 19.9 / 89.6 GB | Unified memory | up to 64k | — |
| Gemma 4 26B-A4B | Runs great | est. 63.7 tok/s51.0–76.5calibrated estimate ±20% | Q4_K_M | 17.9 / 89.6 GB | Unified memory | up to 128k | — |
| DeepSeek R1 Distill Llama 8B | Runs great | est. 61.5 tok/s49.2–73.8calibrated estimate ±20% | Q4_K_M | 6.6 / 89.6 GB | Unified memory | up to 32k |
|
| Kanana 1.5 8B | Runs great | est. 61.5 tok/s49.2–73.8calibrated estimate ±20% | Q4_K_M | 6.6 / 89.6 GB | Unified memory | up to 32k | — |
| Llama 3.1 8B | Runs great | est. 61.5 tok/s49.2–73.8calibrated estimate ±20% | Q4_K_M | 6.6 / 89.6 GB | Unified memory | up to 64k | — |
| Kanana 1.5 15.7B-A3B | Runs great | est. 60.0 tok/s48.0–72.0calibrated estimate ±20% | Q4_K_M | 12.1 / 89.6 GB | Unified memory | up to 32k | — |
| Qwen3.5 9B | Runs great | est. 60.0 tok/s48.0–72.0calibrated estimate ±20% | Q4_K_M | 6.7 / 89.6 GB | Unified memory | up to 256k | — |
| gpt-oss-120b | Runs great | est. 60.0 tok/s48.0–72.0calibrated estimate ±20% | MXFP4 | 64.3 / 89.6 GB | Unified memory | up to 64k |
|
| Qwen3 8B | Runs great | est. 59.1 tok/s47.2–70.9calibrated estimate ±20% | Q4_K_M | 6.8 / 89.6 GB | Unified memory | up to 32k | — |
| Gemma 4 12B | Runs great | est. 48.1 tok/s38.5–57.7calibrated estimate ±20% | Q4_K_M | 8.2 / 89.6 GB | Unified memory | up to 128k | — |
| Gemma 3 12B | Runs great | est. 47.0 tok/s37.6–56.4calibrated estimate ±20% | Q4_K_M | 8.4 / 89.6 GB | Unified memory | up to 128k | — |
| HyperCLOVA X SEED Think 14B | Runs great | est. 36.1 tok/s25.3–47.0theoretical estimate ±30% | Q4_K_M | 10.8 / 89.6 GB | Unified memory | up to 16k |
|
| Qwen3 14B | Runs great | est. 35.6 tok/s28.5–42.7calibrated estimate ±20% | Q4_K_M | 10.9 / 89.6 GB | Unified memory | up to 32k | — |
| Phi-4 | Runs great | est. 34.3 tok/s27.5–41.2calibrated estimate ±20% | Q4_K_M | 11.3 / 89.6 GB | Unified memory | up to 8k |
|
| Qwen3.5 122B-A10B | Runs great | est. 28.5 tok/s22.8–34.2calibrated estimate ±20% | Q4_K_M | 79.0 / 89.6 GB | Unified memory | up to 64k | — |
| Mistral Small 3.2 24B | Runs great | est. 23.5 tok/s18.8–28.2calibrated estimate ±20% | Q4_K_M | 16.2 / 89.6 GB | Unified memory | up to 16k | — |
| Gemma 3 27B | Runs great | est. 21.4 tok/s17.1–25.7calibrated estimate ±20% | Q4_K_M | 17.8 / 89.6 GB | Unified memory | up to 16k | — |
| Qwen3.5 27B | Runs great | est. 20.9 tok/s16.7–25.1calibrated estimate ±20% | Q4_K_M | 18.2 / 89.6 GB | Unified memory | up to 16k | — |
| Solar Open 100B | Runs great | est. 20.7 tok/s16.6–24.9calibrated estimate ±20% | Q4_K_M | 64.4 / 89.6 GB | Unified memory | up to 8k | — |
| EXAONE 4.0 32B | Runs well | est. 18.5 tok/s14.8–22.2calibrated estimate ±20% | Q4_K_M | 20.5 / 89.6 GB | Unified memory | up to 128k |
|
| EXAONE 4.5 33B | Runs well | est. 17.9 tok/s14.3–21.5calibrated estimate ±20% | Q4_K_M | 21.2 / 89.6 GB | Unified memory | up to 128k |
|
| Qwen3 32B | Runs well | est. 16.8 tok/s13.5–20.2calibrated estimate ±20% | Q4_K_M | 22.5 / 89.6 GB | Unified memory | up to 32k |
|
| DeepSeek R1 Distill Qwen 32B | Runs well | est. 16.7 tok/s13.4–20.1calibrated estimate ±20% | Q4_K_M | 22.6 / 89.6 GB | Unified memory | up to 32k |
|
| Llama 3.3 70B | Runs well | est. 8.1 tok/s6.5–9.8calibrated estimate ±20% | Q4_K_M | 45.8 / 89.6 GB | Unified memory | up to 8k | — |
| Solar Open 2 250B | Won't runTry Q2_K (heavy quality loss): Runs slowly | est. 30.1 tok/s21.1–39.1theoretical estimate ±30% | Q2_K | 95.8 GB RAM | CPU only | up to 256k |
|
Reasoning models spend extra tokens thinking, so their speed thresholds are 1.5× stricter (30 / 12 / 3 tok/s).
Verdict by memory size
| Model | 48 GB | 64 GB | 128 GB |
|---|---|---|---|
| EXAONE 4.0 1.2B | Runs great | Runs great | Runs great |
| HyperCLOVA X SEED 1.5B | Runs great | Runs great | Runs great |
| Qwen3.5 35B-A3B | Runs great | Runs great | Runs great |
| gpt-oss-20b | Runs great | Runs great | Runs great |
| Qwen3 30B-A3B (2507) | Runs great | Runs great | Runs great |
| Gemma 4 26B-A4B | Runs great | Runs great | Runs great |
| DeepSeek R1 Distill Llama 8B | Runs great | Runs great | Runs great |
| Kanana 1.5 8B | Runs great | Runs great | Runs great |
| Llama 3.1 8B | Runs great | Runs great | Runs great |
| Kanana 1.5 15.7B-A3B | Runs great | Runs great | Runs great |
| Qwen3.5 9B | Runs great | Runs great | Runs great |
| gpt-oss-120b | Won't run | Won't run | Runs great |
| Qwen3 8B | Runs great | Runs great | Runs great |
| Gemma 4 12B | Runs great | Runs great | Runs great |
| Gemma 3 12B | Runs great | Runs great | Runs great |
| HyperCLOVA X SEED Think 14B | Runs great | Runs great | Runs great |
| Qwen3 14B | Runs great | Runs great | Runs great |
| Phi-4 | Runs great | Runs great | Runs great |
| Qwen3.5 122B-A10B | Won't run | Won't run | Runs great |
| Mistral Small 3.2 24B | Runs great | Runs great | Runs great |
| Gemma 3 27B | Runs great | Runs great | Runs great |
| Qwen3.5 27B | Runs great | Runs great | Runs great |
| Solar Open 100B | Won't run | Won't run | Runs great |
| EXAONE 4.0 32B | Runs well | Runs well | Runs well |
| EXAONE 4.5 33B | Runs well | Runs well | Runs well |
| Qwen3 32B | Runs well | Runs well | Runs well |
| DeepSeek R1 Distill Qwen 32B | Runs well | Runs well | Runs well |
| Llama 3.3 70B | Won't run | Runs slowly | Runs well |
| Solar Open 2 250B | Won't run | Won't run | Won't run |
Measured results on the Apple M5 Max (40-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 (40-core GPU)?
Qwen3.5 122B-A10B at Q4_K_M (a 78.26 GB file) fits entirely in 128 GB of unified memory with 8k context, at est. 28.5 tok/s (22.8–34.2, calibrated estimate ±20%).
How many local LLMs run well on the Apple M5 Max (40-core GPU)?
At Q4_K_M with 8k context, with 128 GB of memory, out of 29 tracked models: 23 run great, 5 run well, 0 run slowly and 1 won't run.
Can the Apple M5 Max (40-core GPU) run a 70B model like Llama 3.3 70B?
Runs well — Unified memory, Q4_K_M: est. 8.1 tok/s (6.5–9.8, calibrated estimate ±20%).
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.