Which local LLMs can the Apple M5 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
- Memory bandwidth
- 1,228.0 GB/s
- FP16 compute
- 33.2 TFLOPS
- Launch year
- 2026
- Macs with this chip
- Mac Studio: 96 GB · 256 GB
- Also written as
- M5 Ultra
Verdicts at a glance
At Q4_K_M (or the closest available quant) with 8k context.
- 96 GBRuns great51 modelsRuns well1 modelRuns slowly3 modelsWon't run3 models
- 256 GBRuns great56 modelsRuns well1 modelRuns slowly1 modelWon'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 M5 Ultra
Scroll sideways to see every column.
| Model | Verdict | Speed | Quant | Memory | Runs as | Context | Notes |
|---|---|---|---|---|---|---|---|
| MiniCPM5 1B | Runs great | est. 785.3 tok/s628.2–942.4calibrated estimate ±20% | Q4_K_M | 1.5 / 179.2 GB | Unified memory | up to 128k |
|
| Llama 3.2 1B | Runs great | est. 649.1 tok/s519.2–778.9calibrated estimate ±20% | Q4_K_M | 1.7 / 179.2 GB | Unified memory | up to 128k | — |
| EXAONE 4.0 1.2B | Runs great | est. 533.0 tok/s426.4–639.6calibrated estimate ±20% | Q4_K_M | 1.9 / 179.2 GB | Unified memory | up to 64k |
|
| LFM2.5 2.6B | Runs great | est. 388.0 tok/s271.6–504.3theoretical estimate ±30% | Q4_K_M | 2.4 / 179.2 GB | Unified memory | up to 128k |
|
| HyperCLOVA X SEED 1.5B | Runs great | est. 385.6 tok/s308.5–462.7calibrated estimate ±20% | Q4_K_M | 2.4 / 179.2 GB | Unified memory | up to 16k | — |
| MiniCPM5 2B | Runs great | est. 366.0 tok/s292.8–439.2calibrated estimate ±20% | Q4_K_M | 2.5 / 179.2 GB | Unified memory | up to 128k |
|
| Ling-3.0 Tiny | Runs great | est. 264.1 tok/s184.9–343.3theoretical estimate ±30% | Q4_K_M | 5.5 / 179.2 GB | Unified memory | up to 128k |
|
| Gemma 4 E2B | Runs great | est. 252.2 tok/s176.5–327.9theoretical estimate ±30% | Q4_K_M | 3.7 / 179.2 GB | Unified memory | up to 128k |
|
| Spark-X2.5 4B | Runs great | est. 241.2 tok/s193.0–289.4calibrated estimate ±20% | Q4_K_M | 3.5 / 179.2 GB | Unified memory | up to 256k |
|
| Llama 3.2 3B | Runs great | est. 236.5 tok/s189.2–283.8calibrated estimate ±20% | Q4_K_M | 3.5 / 179.2 GB | Unified memory | up to 128k | — |
| Qwen3.5 4B | Runs great | est. 225.9 tok/s180.7–271.1calibrated estimate ±20% | Q4_K_M | 3.7 / 179.2 GB | Unified memory | up to 256k |
|
| gpt-oss-20b | Runs great | est. 187.4 tok/s149.9–224.9calibrated estimate ±20% | MXFP4 | 12.9 / 179.2 GB | Unified memory | up to 128k |
|
| Mellum2.1 12B-A2.5B Thinking | Runs great | est. 183.0 tok/s146.4–219.6calibrated estimate ±20% | Q4_K_M | 8.8 / 179.2 GB | Unified memory | up to 128k |
|
| Nemotron 3.5 Lightning 30B-A3B | Runs great | est. 168.6 tok/s118.0–219.2theoretical estimate ±30% | Q4_K_M | 26.1 / 179.2 GB | Unified memory | up to 256k |
|
| Gemma 4 E4B | Runs great | est. 159.1 tok/s111.4–206.8theoretical estimate ±30% | Q4_K_M | 5.7 / 179.2 GB | Unified memory | up to 128k |
|
| gpt-oss-120b | Runs great | est. 131.0 tok/s104.8–157.2calibrated estimate ±20% | MXFP4 | 64.3 / 179.2 GB | Unified memory | up to 128k |
|
| MiniCPM-V 4.7 35B-A3B | Runs great | est. 130.8 tok/s104.7–157.0calibrated estimate ±20% | Q4_K_M | 21.9 / 179.2 GB | Unified memory | up to 256k |
|
| Kanana 1.5 15.7B-A3B | Runs great | est. 128.1 tok/s102.5–153.7calibrated estimate ±20% | Q4_K_M | 12.1 / 179.2 GB | Unified memory | up to 32k | — |
| Qwen3.6 35B-A3B | Runs great | est. 125.7 tok/s100.5–150.8calibrated estimate ±20% | Q4_K_M | 23.0 / 179.2 GB | Unified memory | up to 256k |
|
| Qwen3.5 35B-A3B | Runs great | est. 116.8 tok/s93.5–140.2calibrated estimate ±20% | Q4_K_M | 23.4 / 179.2 GB | Unified memory | up to 256k |
|
| DeepSeek R1 Distill Llama 8B | Runs great | est. 116.8 tok/s93.4–140.1calibrated estimate ±20% | Q4_K_M | 6.6 / 179.2 GB | Unified memory | up to 128k |
|
| Kanana 1.5 8B | Runs great | est. 116.8 tok/s93.4–140.1calibrated estimate ±20% | Q4_K_M | 6.6 / 179.2 GB | Unified memory | up to 32k | — |
| Llama 3.1 8B | Runs great | est. 116.8 tok/s93.4–140.1calibrated estimate ±20% | Q4_K_M | 6.6 / 179.2 GB | Unified memory | up to 128k | — |
| Gemma 4 26B-A4B | Runs great | est. 115.8 tok/s92.6–138.9calibrated estimate ±20% | Q4_K_M | 17.7 / 179.2 GB | Unified memory | up to 256k |
|
| Nanbeige4.2 3B | Runs great | est. 114.3 tok/s80.0–148.5theoretical estimate ±30% | Q4_K_M | 4.7 / 179.2 GB | Unified memory | up to 64k |
|
| Qwen3.5 9B | Runs great | est. 114.0 tok/s91.2–136.8calibrated estimate ±20% | Q4_K_M | 6.7 / 179.2 GB | Unified memory | up to 256k |
|
| DeepSeek R1 0528 Qwen3 8B | Runs great | est. 112.2 tok/s89.8–134.7calibrated estimate ±20% | Q4_K_M | 6.8 / 179.2 GB | Unified memory | up to 64k |
|
| Qwen3 8B | Runs great | est. 112.2 tok/s89.8–134.7calibrated estimate ±20% | Q4_K_M | 6.8 / 179.2 GB | Unified memory | up to 32k |
|
| Kolibri 1 | Runs great | est. 108.1 tok/s86.5–129.8calibrated estimate ±20% | Q4_K_M | 50.5 / 179.2 GB | Unified memory | up to 256k |
|
| Qwen3 30B-A3B (2507) | Runs great | est. 106.6 tok/s85.3–127.9calibrated estimate ±20% | Q4_K_M | 19.9 / 179.2 GB | Unified memory | up to 256k | — |
| Ling-3.0 Flash | Runs great | est. 106.3 tok/s74.4–138.2theoretical estimate ±30% | Q4_K_M | 77.7 / 179.2 GB | Unified memory | up to 256k |
|
| K2-Horizon 7B | Runs great | est. 103.0 tok/s82.4–123.6calibrated estimate ±20% | Q4_K_M | 7.4 / 179.2 GB | Unified memory | up to 64k |
|
| Gemma 4 12B | Runs great | est. 96.5 tok/s77.2–115.8calibrated estimate ±20% | Q4_K_M | 7.8 / 179.2 GB | Unified memory | up to 256k |
|
| Gemma 3 12B | Runs great | est. 89.3 tok/s71.5–107.2calibrated estimate ±20% | Q4_K_M | 8.4 / 179.2 GB | Unified memory | up to 128k | — |
| K2-Horizon MoVA 36B-A4B | Runs great | est. 82.3 tok/s57.6–106.9theoretical estimate ±30% | Q4_K_M | 24.6 / 179.2 GB | Unified memory | up to 64k |
|
| HyperCLOVA X SEED Think 14B | Runs great | est. 68.7 tok/s48.1–89.3theoretical estimate ±30% | Q4_K_M | 10.8 / 179.2 GB | Unified memory | up to 64k |
|
| Qwen3 14B | Runs great | est. 67.7 tok/s54.1–81.2calibrated estimate ±20% | Q4_K_M | 10.9 / 179.2 GB | Unified memory | up to 32k |
|
| DeepSeek R1 Distill Qwen 14B | Runs great | est. 66.0 tok/s52.8–79.2calibrated estimate ±20% | Q4_K_M | 11.2 / 179.2 GB | Unified memory | up to 64k |
|
| Phi-4 | Runs great | est. 65.2 tok/s52.2–78.3calibrated estimate ±20% | Q4_K_M | 11.3 / 179.2 GB | Unified memory | up to 16k | — |
| Laguna S 2.1 | Runs great | est. 61.9 tok/s49.5–74.3calibrated estimate ±20% | Q4_K_M | 68.6 / 179.2 GB | Unified memory | up to 128k |
|
| Qwen3.5 122B-A10B | Runs great | est. 58.6 tok/s46.9–70.4calibrated estimate ±20% | Q4_K_M | 79.0 / 179.2 GB | Unified memory | up to 256k |
|
| Solar Open 100B | Runs great | est. 56.5 tok/s45.2–67.8calibrated estimate ±20% | Q4_K_M | 64.4 / 179.2 GB | Unified memory | up to 32k |
|
| DeepSeek V4 Flash 0731 | Runs great | est. 52.5 tok/s36.7–68.2theoretical estimate ±30% | MXFP4 | 155.6 / 179.2 GB | Unified memory | up to 256k |
|
| Solar Open 2 250B | Runs great | est. 50.3 tok/s40.3–60.4calibrated estimate ±20% | IQ4_XS | 137.2 / 179.2 GB | Unified memory | up to 128k |
|
| Mistral Small 3.2 24B | Runs great | est. 44.7 tok/s35.7–53.6calibrated estimate ±20% | Q4_K_M | 16.2 / 179.2 GB | Unified memory | up to 64k | — |
| Gemma 3 27B | Runs great | est. 40.6 tok/s32.5–48.8calibrated estimate ±20% | Q4_K_M | 17.8 / 179.2 GB | Unified memory | up to 128k | — |
| Qwen3.6 27B | Runs great | est. 40.3 tok/s32.3–48.4calibrated estimate ±20% | Q4_K_M | 17.9 / 179.2 GB | Unified memory | up to 64k |
|
| Muse Glimmer 30B | Runs great | est. 40.2 tok/s32.1–48.2calibrated estimate ±20% | Q4_K_M | 18.0 / 179.2 GB | Unified memory | up to 128k |
|
| Qwen3.5 27B | Runs great | est. 39.7 tok/s31.7–47.6calibrated estimate ±20% | Q4_K_M | 18.2 / 179.2 GB | Unified memory | up to 64k |
|
| Qwen3.8 27B | Runs great | est. 38.9 tok/s31.1–46.7calibrated estimate ±20% | Q4_K_M | 18.6 / 179.2 GB | Unified memory | up to 64k |
|
| Gemma 4 31B | Runs great | est. 36.9 tok/s29.5–44.2calibrated estimate ±20% | Q4_K_M | 19.6 / 179.2 GB | Unified memory | up to 32k |
|
| Granite 4.2 30B | Runs great | est. 35.2 tok/s28.2–42.3calibrated estimate ±20% | Q4_K_M | 20.4 / 179.2 GB | Unified memory | up to 16k |
|
| EXAONE 4.0 32B | Runs great | est. 35.2 tok/s28.2–42.3calibrated estimate ±20% | Q4_K_M | 20.5 / 179.2 GB | Unified memory | up to 32k |
|
| EXAONE 4.5 33B | Runs great | est. 34.0 tok/s27.2–40.8calibrated estimate ±20% | Q4_K_M | 21.2 / 179.2 GB | Unified memory | up to 32k |
|
| Qwen3 32B | Runs great | est. 32.0 tok/s25.6–38.3calibrated estimate ±20% | Q4_K_M | 22.5 / 179.2 GB | Unified memory | up to 8k |
|
| DeepSeek R1 Distill Qwen 32B | Runs great | est. 31.8 tok/s25.5–38.2calibrated estimate ±20% | Q4_K_M | 22.6 / 179.2 GB | Unified memory | up to 8k |
|
| Llama 3.3 70B | Runs well | est. 15.5 tok/s12.4–18.6calibrated estimate ±20% | Q4_K_M | 45.8 / 179.2 GB | Unified memory | up to 128k |
|
| GLM-5.3 Flash | Runs slowly | est. 39.5 tok/s27.6–51.3theoretical estimate ±30% | Q4_K_M | 181.5 GB RAM | CPU only | up to 256k |
|
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 |
|---|---|---|
| MiniCPM5 1B | Runs great | Runs great |
| Llama 3.2 1B | Runs great | Runs great |
| EXAONE 4.0 1.2B | Runs great | Runs great |
| LFM2.5 2.6B | Runs great | Runs great |
| HyperCLOVA X SEED 1.5B | Runs great | Runs great |
| MiniCPM5 2B | Runs great | Runs great |
| Ling-3.0 Tiny | Runs great | Runs great |
| Gemma 4 E2B | Runs great | Runs great |
| Spark-X2.5 4B | Runs great | Runs great |
| Llama 3.2 3B | Runs great | Runs great |
| Qwen3.5 4B | Runs great | Runs great |
| gpt-oss-20b | Runs great | Runs great |
| Mellum2.1 12B-A2.5B Thinking | Runs great | Runs great |
| Nemotron 3.5 Lightning 30B-A3B | Runs great | Runs great |
| Gemma 4 E4B | Runs great | Runs great |
| gpt-oss-120b | Runs great | Runs great |
| MiniCPM-V 4.7 35B-A3B | Runs great | Runs great |
| Kanana 1.5 15.7B-A3B | Runs great | Runs great |
| Qwen3.6 35B-A3B | Runs great | Runs great |
| Qwen3.5 35B-A3B | 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 |
| Gemma 4 26B-A4B | Runs great | Runs great |
| Nanbeige4.2 3B | Runs great | Runs great |
| Qwen3.5 9B | Runs great | Runs great |
| DeepSeek R1 0528 Qwen3 8B | Runs great | Runs great |
| Qwen3 8B | Runs great | Runs great |
| Kolibri 1 | Runs great | Runs great |
| Qwen3 30B-A3B (2507) | Runs great | Runs great |
| Ling-3.0 Flash | Runs slowly | Runs great |
| K2-Horizon 7B | Runs great | Runs great |
| Gemma 4 12B | Runs great | Runs great |
| Gemma 3 12B | Runs great | Runs great |
| K2-Horizon MoVA 36B-A4B | Runs great | Runs great |
| HyperCLOVA X SEED Think 14B | Runs great | Runs great |
| Qwen3 14B | Runs great | Runs great |
| DeepSeek R1 Distill Qwen 14B | Runs great | Runs great |
| Phi-4 | Runs great | Runs great |
| Laguna S 2.1 | Runs slowly | Runs great |
| Qwen3.5 122B-A10B | Runs slowly | Runs great |
| Solar Open 100B | Runs great | Runs great |
| DeepSeek V4 Flash 0731 | Won't run | Runs great |
| Solar Open 2 250B | Won't run | Runs great |
| Mistral Small 3.2 24B | Runs great | Runs great |
| Gemma 3 27B | Runs great | Runs great |
| Qwen3.6 27B | Runs great | Runs great |
| Muse Glimmer 30B | Runs great | Runs great |
| Qwen3.5 27B | Runs great | Runs great |
| Qwen3.8 27B | Runs great | Runs great |
| Gemma 4 31B | Runs great | Runs great |
| Granite 4.2 30B | Runs great | Runs great |
| EXAONE 4.0 32B | Runs great | Runs great |
| EXAONE 4.5 33B | Runs great | Runs great |
| Qwen3 32B | Runs great | Runs great |
| DeepSeek R1 Distill Qwen 32B | Runs great | Runs great |
| Llama 3.3 70B | Runs well | Runs well |
| GLM-5.3 Flash | Won't run | Runs slowly |
Measured results on the Apple M5 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 |
|---|---|---|---|---|---|---|---|---|
| llama-2-7b | Q4_0 | llama.cpp | 512 | 4,945.0 | 179.1 | Metal, FA | github.com | 2026-09-30 |
Frequently asked questions
What is the largest model that runs entirely on the Apple M5 Ultra?
DeepSeek V4 Flash 0731 at MXFP4 (a 154.99 GB file) fits entirely in 256 GB of unified memory with 8k context, at est. 52.5 tok/s (36.7–68.2, theoretical estimate ±30%).
How many local LLMs run well on the Apple M5 Ultra?
At Q4_K_M with 8k context, with 256 GB of memory, out of 58 tracked models: 56 run great, 1 runs well, 1 runs slowly and 0 won't run.
Can the Apple M5 Ultra run a 70B model like Llama 3.3 70B?
Runs well — Unified memory, Q4_K_M: est. 15.5 tok/s (12.4–18.6, 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.