Which local LLMs can the Apple M3 Ultra (Mac Studio) 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
- 96 GB · 256 GB · 512 GB
- Memory bandwidth
- 819.0 GB/s
- FP16 compute
- 28.3 TFLOPS
- Launch year
- 2025
- Macs with this chip
- Mac Studio: 96 GB · 256 GB · 512 GB
- Also written as
- M3 Ultra
Verdicts at a glance
At Q4_K_M (or the closest available quant) with 8k context.
- 96 GBRuns great41 modelsRuns well12 modelsRuns slowly3 modelsWon't run3 models
- 256 GBRuns great46 modelsRuns well12 modelsRuns slowly1 modelWon't run0 models
- 512 GBRuns great46 modelsRuns well13 modelsRuns slowly0 modelsWon't run0 models
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 M3 Ultra
Scroll sideways to see every column.
| Model | Verdict | Speed | Quant | Memory | Runs as | Context | Notes |
|---|---|---|---|---|---|---|---|
| MiniCPM5 1B | Runs great | est. 523.7 tok/s419.0–628.5calibrated estimate ±20% | Q4_K_M | 1.5 / 358.4 GB | Unified memory | up to 128k |
|
| Llama 3.2 1B | Runs great | est. 432.9 tok/s346.3–519.5calibrated estimate ±20% | Q4_K_M | 1.7 / 358.4 GB | Unified memory | up to 128k | — |
| EXAONE 4.0 1.2B | Runs great | est. 355.5 tok/s284.4–426.6calibrated estimate ±20% | Q4_K_M | 1.9 / 358.4 GB | Unified memory | up to 64k |
|
| LFM2.5 2.6B | Runs great | est. 258.7 tok/s181.1–336.4theoretical estimate ±30% | Q4_K_M | 2.4 / 358.4 GB | Unified memory | up to 128k |
|
| HyperCLOVA X SEED 1.5B | Runs great | est. 257.2 tok/s205.7–308.6calibrated estimate ±20% | Q4_K_M | 2.4 / 358.4 GB | Unified memory | up to 16k | — |
| MiniCPM5 2B | Runs great | est. 244.1 tok/s195.3–292.9calibrated estimate ±20% | Q4_K_M | 2.5 / 358.4 GB | Unified memory | up to 128k |
|
| Ling-3.0 Tiny | Runs great | est. 229.2 tok/s160.4–298.0theoretical estimate ±30% | Q4_K_M | 5.5 / 358.4 GB | Unified memory | up to 128k |
|
| Gemma 4 E2B | Runs great | est. 197.2 tok/s138.0–256.4theoretical estimate ±30% | Q4_K_M | 3.7 / 358.4 GB | Unified memory | up to 128k |
|
| Spark-X2.5 4B | Runs great | est. 160.9 tok/s128.7–193.0calibrated estimate ±20% | Q4_K_M | 3.5 / 358.4 GB | Unified memory | up to 256k |
|
| Llama 3.2 3B | Runs great | est. 157.7 tok/s126.2–189.3calibrated estimate ±20% | Q4_K_M | 3.5 / 358.4 GB | Unified memory | up to 128k | — |
| Mellum2.1 12B-A2.5B Thinking | Runs great | est. 151.7 tok/s121.3–182.0calibrated estimate ±20% | Q4_K_M | 8.8 / 358.4 GB | Unified memory | up to 128k |
|
| Qwen3.5 4B | Runs great | est. 150.7 tok/s120.5–180.8calibrated estimate ±20% | Q4_K_M | 3.7 / 358.4 GB | Unified memory | up to 256k |
|
| Nemotron 3.5 Lightning 30B-A3B | Runs great | est. 126.5 tok/s88.5–164.4theoretical estimate ±30% | Q4_K_M | 26.1 / 358.4 GB | Unified memory | up to 256k |
|
| Gemma 4 E4B | Runs great | est. 117.0 tok/s81.9–152.1theoretical estimate ±30% | Q4_K_M | 5.7 / 358.4 GB | Unified memory | up to 128k |
|
| gpt-oss-20b | Runs great | 115.5 tok/s8k estimate 136.7 tok/s (109.4–164.1, calibrated estimate ±20%)measured (1 run, 2k context) | MXFP4 | 12.9 / 358.4 GB | Unified memory | up to 128k |
|
| MiniCPM-V 4.7 35B-A3B | Runs great | est. 109.0 tok/s87.2–130.8calibrated estimate ±20% | Q4_K_M | 21.9 / 358.4 GB | Unified memory | up to 256k |
|
| Qwen3.6 35B-A3B | Runs great | est. 103.7 tok/s82.9–124.4calibrated estimate ±20% | Q4_K_M | 23.0 / 358.4 GB | Unified memory | up to 256k |
|
| Kanana 1.5 15.7B-A3B | Runs great | est. 101.3 tok/s81.0–121.5calibrated estimate ±20% | Q4_K_M | 12.1 / 358.4 GB | Unified memory | up to 32k | — |
| gpt-oss-120b | Runs great | est. 96.0 tok/s76.8–115.2calibrated estimate ±20% | MXFP4 | 64.3 / 358.4 GB | Unified memory | up to 128k |
|
| Qwen3.5 35B-A3B | Runs great | est. 94.8 tok/s75.8–113.8calibrated estimate ±20% | Q4_K_M | 23.4 / 358.4 GB | Unified memory | up to 256k |
|
| Gemma 4 26B-A4B | Runs great | est. 89.0 tok/s71.2–106.8calibrated estimate ±20% | Q4_K_M | 17.7 / 358.4 GB | Unified memory | up to 256k |
|
| Qwen3 30B-A3B (2507) | Runs great | est. 88.3 tok/s70.7–106.0calibrated estimate ±20% | Q4_K_M | 19.9 / 358.4 GB | Unified memory | up to 128k | — |
| Kolibri 1 | Runs great | est. 85.3 tok/s68.2–102.4calibrated estimate ±20% | Q4_K_M | 50.5 / 358.4 GB | Unified memory | up to 256k |
|
| Ling-3.0 Flash | Runs great | est. 84.6 tok/s59.2–110.0theoretical estimate ±30% | Q4_K_M | 77.7 / 358.4 GB | Unified memory | up to 256k |
|
| DeepSeek R1 Distill Llama 8B | Runs great | est. 77.9 tok/s62.3–93.5calibrated estimate ±20% | Q4_K_M | 6.6 / 358.4 GB | Unified memory | up to 64k |
|
| Kanana 1.5 8B | Runs great | est. 77.9 tok/s62.3–93.5calibrated estimate ±20% | Q4_K_M | 6.6 / 358.4 GB | Unified memory | up to 32k | — |
| Llama 3.1 8B | Runs great | est. 77.9 tok/s62.3–93.5calibrated estimate ±20% | Q4_K_M | 6.6 / 358.4 GB | Unified memory | up to 128k | — |
| Nanbeige4.2 3B | Runs great | est. 76.2 tok/s53.3–99.1theoretical estimate ±30% | Q4_K_M | 4.7 / 358.4 GB | Unified memory | up to 32k |
|
| Qwen3.5 9B | Runs great | est. 76.1 tok/s60.8–91.3calibrated estimate ±20% | Q4_K_M | 6.7 / 358.4 GB | Unified memory | up to 256k |
|
| DeepSeek R1 0528 Qwen3 8B | Runs great | est. 74.8 tok/s59.9–89.8calibrated estimate ±20% | Q4_K_M | 6.8 / 358.4 GB | Unified memory | up to 64k |
|
| Qwen3 8B | Runs great | est. 74.8 tok/s59.9–89.8calibrated estimate ±20% | Q4_K_M | 6.8 / 358.4 GB | Unified memory | up to 32k |
|
| K2-Horizon 7B | Runs great | est. 68.7 tok/s54.9–82.4calibrated estimate ±20% | Q4_K_M | 7.4 / 358.4 GB | Unified memory | up to 64k |
|
| Gemma 4 12B | Runs great | est. 64.4 tok/s51.5–77.2calibrated estimate ±20% | Q4_K_M | 7.8 / 358.4 GB | Unified memory | up to 256k |
|
| K2-Horizon MoVA 36B-A4B | Runs great | est. 63.9 tok/s44.7–83.0theoretical estimate ±30% | Q4_K_M | 24.6 / 358.4 GB | Unified memory | up to 32k |
|
| Gemma 3 12B | Runs great | est. 59.6 tok/s47.7–71.5calibrated estimate ±20% | Q4_K_M | 8.4 / 358.4 GB | Unified memory | up to 128k | — |
| Laguna S 2.1 | Runs great | est. 48.1 tok/s38.5–57.7calibrated estimate ±20% | Q4_K_M | 68.6 / 358.4 GB | Unified memory | up to 64k |
|
| HyperCLOVA X SEED Think 14B | Runs great | est. 45.8 tok/s32.1–59.5theoretical estimate ±30% | Q4_K_M | 10.8 / 358.4 GB | Unified memory | up to 32k |
|
| Qwen3 14B | Runs great | est. 45.1 tok/s36.1–54.2calibrated estimate ±20% | Q4_K_M | 10.9 / 358.4 GB | Unified memory | up to 32k |
|
| DeepSeek R1 Distill Qwen 14B | Runs great | est. 44.0 tok/s35.2–52.8calibrated estimate ±20% | Q4_K_M | 11.2 / 358.4 GB | Unified memory | up to 32k |
|
| Qwen3.5 122B-A10B | Runs great | est. 43.8 tok/s35.0–52.6calibrated estimate ±20% | Q4_K_M | 79.0 / 358.4 GB | Unified memory | up to 128k |
|
| Phi-4 | Runs great | est. 43.5 tok/s34.8–52.2calibrated estimate ±20% | Q4_K_M | 11.3 / 358.4 GB | Unified memory | up to 16k | — |
| Solar Open 100B | Runs great | est. 42.0 tok/s33.6–50.4calibrated estimate ±20% | Q4_K_M | 64.4 / 358.4 GB | Unified memory | up to 16k |
|
| DeepSeek V4 Flash 0731 | Runs great | est. 37.4 tok/s26.2–48.6theoretical estimate ±30% | MXFP4 | 155.6 / 358.4 GB | Unified memory | up to 32k |
|
| Solar Open 2 250B | Runs great | est. 37.0 tok/s29.6–44.4calibrated estimate ±20% | IQ4_XS | 137.2 / 358.4 GB | Unified memory | up to 64k |
|
| Mistral Small 3.2 24B | Runs great | est. 29.8 tok/s23.8–35.7calibrated estimate ±20% | Q4_K_M | 16.2 / 358.4 GB | Unified memory | up to 32k | — |
| Gemma 3 27B | Runs great | est. 27.1 tok/s21.7–32.5calibrated estimate ±20% | Q4_K_M | 17.8 / 358.4 GB | Unified memory | up to 64k | — |
| GLM-5.3 Flash | Runs well | est. 28.1 tok/s19.7–36.5theoretical estimate ±30% | Q4_K_M | 182.1 / 358.4 GB | Unified memory | up to 256k |
|
| Qwen3.6 27B | Runs well | est. 26.9 tok/s21.5–32.3calibrated estimate ±20% | Q4_K_M | 17.9 / 358.4 GB | Unified memory | up to 256k |
|
| Muse Glimmer 30B | Runs well | est. 26.8 tok/s21.4–32.2calibrated estimate ±20% | Q4_K_M | 18.0 / 358.4 GB | Unified memory | up to 128k |
|
| Qwen3 32B | Runs well | 26.7 tok/s8k estimate 21.3 tok/s (17.0–25.6, calibrated estimate ±20%)measured (1 run, 512 context) | Q4_K_M | 22.5 / 358.4 GB | Unified memory | up to 32k |
|
| Qwen3.5 27B | Runs well | est. 26.5 tok/s21.2–31.7calibrated estimate ±20% | Q4_K_M | 18.2 / 358.4 GB | Unified memory | up to 256k |
|
| Qwen3.8 27B | Runs well | est. 26.0 tok/s20.8–31.2calibrated estimate ±20% | Q4_K_M | 18.6 / 358.4 GB | Unified memory | up to 256k |
|
| Gemma 4 31B | Runs well | est. 24.6 tok/s19.7–29.5calibrated estimate ±20% | Q4_K_M | 19.6 / 358.4 GB | Unified memory | up to 128k |
|
| Granite 4.2 30B | Runs well | est. 23.5 tok/s18.8–28.2calibrated estimate ±20% | Q4_K_M | 20.4 / 358.4 GB | Unified memory | up to 64k |
|
| EXAONE 4.0 32B | Runs well | est. 23.5 tok/s18.8–28.2calibrated estimate ±20% | Q4_K_M | 20.5 / 358.4 GB | Unified memory | up to 128k |
|
| EXAONE 4.5 33B | Runs well | est. 22.7 tok/s18.1–27.2calibrated estimate ±20% | Q4_K_M | 21.2 / 358.4 GB | Unified memory | up to 256k |
|
| DeepSeek R1 Distill Qwen 32B | Runs well | est. 21.2 tok/s17.0–25.5calibrated estimate ±20% | Q4_K_M | 22.6 / 358.4 GB | Unified memory | up to 64k |
|
| Nanbeige4.3 33B | Runs well | est. 20.7 tok/s16.6–24.9calibrated estimate ±20% | Q4_K_M | 23.1 / 358.4 GB | Unified memory | up to 64k |
|
| Llama 3.3 70B | Runs well | 14.4 tok/s8k estimate 10.3 tok/s (8.3–12.4, calibrated estimate ±20%)measured (1 run, 512 context) | Q4_K_M | 45.8 / 358.4 GB | Unified memory | up to 32k | — |
Models with a thinking mode spend extra tokens thinking, so their speed thresholds are 1.5× stricter (30 / 12 / 3 tok/s). With thinking turned off, a model may do one grade better.
Verdict by memory size
| Model | 96 GB | 256 GB | 512 GB |
|---|---|---|---|
| MiniCPM5 1B | Runs great | Runs great | Runs great |
| Llama 3.2 1B | Runs great | Runs great | Runs great |
| EXAONE 4.0 1.2B | Runs great | Runs great | Runs great |
| LFM2.5 2.6B | Runs great | Runs great | Runs great |
| HyperCLOVA X SEED 1.5B | Runs great | Runs great | Runs great |
| MiniCPM5 2B | Runs great | Runs great | Runs great |
| Ling-3.0 Tiny | Runs great | Runs great | Runs great |
| Gemma 4 E2B | Runs great | Runs great | Runs great |
| Spark-X2.5 4B | Runs great | Runs great | Runs great |
| Llama 3.2 3B | Runs great | Runs great | Runs great |
| Mellum2.1 12B-A2.5B Thinking | Runs great | Runs great | Runs great |
| Qwen3.5 4B | Runs great | Runs great | Runs great |
| Nemotron 3.5 Lightning 30B-A3B | Runs great | Runs great | Runs great |
| Gemma 4 E4B | Runs great | Runs great | Runs great |
| gpt-oss-20b | Runs great | Runs great | Runs great |
| MiniCPM-V 4.7 35B-A3B | Runs great | Runs great | Runs great |
| Qwen3.6 35B-A3B | Runs great | Runs great | Runs great |
| Kanana 1.5 15.7B-A3B | Runs great | Runs great | Runs great |
| gpt-oss-120b | Runs great | Runs great | Runs great |
| Qwen3.5 35B-A3B | Runs great | Runs great | Runs great |
| Gemma 4 26B-A4B | Runs great | Runs great | Runs great |
| Qwen3 30B-A3B (2507) | Runs great | Runs great | Runs great |
| Kolibri 1 | Runs great | Runs great | Runs great |
| Ling-3.0 Flash | Runs slowly | 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 |
| Nanbeige4.2 3B | Runs great | Runs great | Runs great |
| Qwen3.5 9B | Runs great | Runs great | Runs great |
| DeepSeek R1 0528 Qwen3 8B | Runs great | Runs great | Runs great |
| Qwen3 8B | Runs great | Runs great | Runs great |
| K2-Horizon 7B | Runs great | Runs great | Runs great |
| Gemma 4 12B | Runs great | Runs great | Runs great |
| K2-Horizon MoVA 36B-A4B | Runs great | Runs great | Runs great |
| Gemma 3 12B | Runs great | Runs great | Runs great |
| Laguna S 2.1 | Runs slowly | Runs great | Runs great |
| HyperCLOVA X SEED Think 14B | Runs great | Runs great | Runs great |
| Qwen3 14B | Runs great | Runs great | Runs great |
| DeepSeek R1 Distill Qwen 14B | Runs great | Runs great | Runs great |
| Qwen3.5 122B-A10B | Runs slowly | Runs great | Runs great |
| Phi-4 | Runs great | Runs great | Runs great |
| Solar Open 100B | Runs great | Runs great | Runs great |
| DeepSeek V4 Flash 0731 | Won't run | Runs great | Runs great |
| Solar Open 2 250B | Won't run | Runs great | Runs great |
| Mistral Small 3.2 24B | Runs great | Runs great | Runs great |
| Gemma 3 27B | Runs great | Runs great | Runs great |
| GLM-5.3 Flash | Won't run | Runs slowly | Runs well |
| Qwen3.6 27B | Runs well | Runs well | Runs well |
| Muse Glimmer 30B | Runs well | Runs well | Runs well |
| Qwen3 32B | Runs well | Runs well | Runs well |
| Qwen3.5 27B | Runs well | Runs well | Runs well |
| Qwen3.8 27B | Runs well | Runs well | Runs well |
| Gemma 4 31B | Runs well | Runs well | Runs well |
| Granite 4.2 30B | Runs well | Runs well | Runs well |
| EXAONE 4.0 32B | Runs well | Runs well | Runs well |
| EXAONE 4.5 33B | Runs well | Runs well | Runs well |
| DeepSeek R1 Distill Qwen 32B | Runs well | Runs well | Runs well |
| Nanbeige4.3 33B | Runs well | Runs well | Runs well |
| Llama 3.3 70B | Runs well | Runs well | Runs well |
Measured results on the Apple M3 Ultra
Public benchmarks we calibrate against. Their conditions (context, backend, flags) can differ from the estimates above.
| Model | Quant | Backend | Context | Prompt (tok/s) | Generation (tok/s) | Flags | Source | Measured |
|---|---|---|---|---|---|---|---|---|
| gpt-oss-20b | MXFP4 | llama.cpp | 2k | 2,816.0 | 115.5 | — | github.com | 2025-08-28 |
| llama-2-7b | Q4_0 | llama.cpp | 512 | 1,471.0 | 92.1 | Metal, b1550 (2023) | github.com | 2025-03-23 |
| Qwen3.5 27B | Q8_0 | llama.cpp | 32k | — | 19.6 | — | github.com | 2026-09-23 |
| Llama 3.3 70B | Q4_K_M | llama.cpp | 512 | 162.0 | 14.4 | Metal | github.com | 2025-12-10 |
| Qwen3 32B | Q4_K_M | llama.cpp | 512 | 335.0 | 26.7 | Metal | github.com | 2025-12-08 |
Frequently asked questions
What is the largest model that runs entirely on the Apple M3 Ultra?
GLM-5.3 Flash at Q4_K_M (a 181.37 GB file) fits entirely in 512 GB of unified memory with 8k context, at est. 28.1 tok/s (19.7–36.5, theoretical estimate ±30%).
How many local LLMs run well on the Apple M3 Ultra?
At Q4_K_M with 8k context, with 512 GB of memory, out of 59 tracked models: 46 run great, 13 run well, 0 run slowly and 0 won't run.
Can the Apple M3 Ultra run a 70B model like Llama 3.3 70B?
Runs well — Unified memory, Q4_K_M: 14.4 tok/s measured at 512 context (1 run; 8k estimate 10.3 tok/s, 8.3–12.4, 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.