Mac M2 Ultra 192G — what LLMs can it run?
101 of 110 indexed models fit comfortably in 192GB at 4K context, each at the highest-quality quant that still leaves headroom.
The short answer for Mac M2 Ultra 192G
192 GB, 800 GB/s Unified LPDDR5. 101 of 110 models in this index fit comfortably at 4K context; 101 load at all.
- Biggest that fits
- Qwen3 235B-A22B Instruct — 119.6 GB at Q3_K_M — 24.4 GB spare at 4K, so longer context comes out of a thin margin.
- Room to grow
- DBRX Instruct — 84.3 GB at Q4_K_M — under 60% of the card, which leaves 59.7 GB for a long context window or a second process.
- Speed ceiling
- Llama 3.1 8B Instruct at Q4_K_M reads 4.6 GB of weights per token, so 800 GB/s puts a hard ceiling near 173 tok/s. That is arithmetic on two published numbers, not a benchmark — real throughput lands below it. This index has 1 measured run on this card; see the table below for what they actually returned.
- The wall you will hit
- The GPU's share of unified memory, not the 192 GB on the box. macOS reserves part of the pool for the system and caps what a single process may wire down, so the practical budget is meaningfully below nameplate — the limit is adjustable (`iogpu.wired_limit_mb`) but it is not absent. Bandwidth is 800 GB/s, which is the number that decides tok/s once a model fits.
70B+ · 14
| Model | Quant | Est. VRAM | Headroom | tok/s |
|---|---|---|---|---|
| Qwen3 235B-A22B Instruct235B-A22B | Q3_K_M | 119.6 GB | +24.4 GB | ≤79Ceiling |
| DBRX Instruct132B MoE | Q4_K_M | 84.31 GB | +59.7 GB | ≤39Ceiling |
| Qwen3.8-Flash-Next125B-A6B | Q4_K_M | 79.32 GB | +64.7 GB | ≤232Ceiling |
| Nemotron 3.5 Super VL 120B-A12B120B-A12B | Q4_0 | 75.12 GB | +68.9 GB | ≤118Ceiling |
| GPT-OSS 120B117B MoE | MXFP4 | 66.4 GB | +77.6 GB | ≤304Ceiling |
| Llama 4 Scout 17B (16E)109B MoE | Q4_K_M | 69.88 GB | +74.1 GB | ≤82Ceiling |
| GLM-4.5-Air106B MoE | Q4_K_M | 67.94 GB | +76.1 GB | ≤116Ceiling |
| Llama 3.2 90B Vision Instruct90B | Q4_K_M | 57.5 GB | +86.5 GB | ≤16Ceiling |
| Nemotron 3.5 Bolt 74B-A9B74B-A9B | IQ2_XXS | 23.05 GB | +121 GB | ≤315Ceiling |
| Qwen2.5 72B Instruct72B | Q5_K_M | 55.31 GB | +88.7 GB | ≤16Ceiling |
| Llama 3.1 70B Instruct70B | Q5_K_M | 53.75 GB | +90.3 GB | ≤17Ceiling |
| Llama 3.3 70B Instruct70B | Q5_K_M | 53.75 GB | +90.3 GB | ≤17Ceiling |
| DeepSeek-R1-Distill-Llama-70B70B | Q4_K_M | 46.1 GB | +97.9 GB | ≤20Ceiling |
| Jamba 1.5 Mini52B-A12B | Q4_K_M | 33.01 GB | +111 GB | ≤116Ceiling |
32B · 30
| Model | Quant | Est. VRAM | Headroom | tok/s |
|---|---|---|---|---|
| Kimi Linear 48B-A3B Instruct48B-A3B | Q8_0 | 53.33 GB | +90.7 GB | ≤264Ceiling |
| Mixtral 8x7B Instruct47B MoE | Q4_K_M | 30.13 GB | +113.9 GB | ≤108Ceiling |
| Seed-OSS 36B Instruct36B | Q5_K_M | 27.81 GB | +116.2 GB | ≤33Ceiling |
| Qwen3.6 35B-A3B35B-A3B | Q8_0 | 39.6 GB | +104.4 GB | ≤264Ceiling |
| Command R 35B35B | Q4_K_M | 22.86 GB | +121.1 GB | ≤40Ceiling |
| Yi 1.5 34B Chat34B | Q4_K_M | 22.82 GB | +121.2 GB | ≤40Ceiling |
| Qwen3 32B Instruct32B | Q4_K_M | 21.85 GB | +122.2 GB | 17–22Derived |
| Qwen2.5 32B Instruct32B | Q4_K_M | 21.69 GB | +122.3 GB | 17–23Derived |
| Qwen2.5-Coder 32B Instruct32B | Q4_K_M | 21.69 GB | +122.3 GB | 17–23Derived |
| DeepSeek-R1-Distill-Qwen-32B32B | Q4_K_M | 21.69 GB | +122.3 GB | 17–23Derived |
| Nemotron 3.5 Lightning 30B-A3B30B-A3B | Q8_0 | 35.14 GB | +108.9 GB | ≤264Ceiling |
| Gemma 4 31B IT31B | Q8_0 | 35.72 GB | +108.3 GB | 10–14Derived |
| Qwen3 30B-A3B Instruct30B-A3B | Q5_K_M | 23.04 GB | +121 GB | ≤395Ceiling |
| Qwen3-Coder 30B-A3B Instruct30B-A3B | Q5_K_M | 23.04 GB | +121 GB | ≤395Ceiling |
| Qwen3-VL 30B-A3B Instruct30B-A3B | Q8_0 | 34.28 GB | +109.7 GB | ≤264Ceiling |
| GLM-4.7-Flash30B-A3B | Q8_0 | 33.47 GB | +110.5 GB | ≤264Ceiling |
| Granite 4.2 30B30B | Q8_0 | 33.61 GB | +110.4 GB | 11–14Derived |
| Muse Glimmer 30B30B | Q8_0 | 31.08 GB | +112.9 GB | 11–15Derived |
| Qwen3.8 27B27B | Q4_K_M | 17.47 GB | +126.5 GB | 20–27Derived |
| Qwen3.6 27B27B | Q8_0 | 30.71 GB | +113.3 GB | 12–15Derived |
| Gemma 3 27B IT27B | Q4_K_M | 18.37 GB | +125.6 GB | 20–27Derived |
| Gemma 2 27B Instruct27B | Q5_K_M | 21.76 GB | +122.2 GB | 17–23Derived |
| Gemma 4 26B-A4B IT26B-A4B | Q8_0 | 28.42 GB | +115.6 GB | ≤198Ceiling |
| Mistral Small 24B Instruct24B | Q4_K_M | 15.89 GB | +128.1 GB | 23–31Derived |
| LFM2.5 24B-A2B24B-A2B | Q8_0 | 26.73 GB | +117.3 GB | ≤323Ceiling |
| Devstral Small 1.1 24B24B | Q6_K | 20.91 GB | +123.1 GB | 17–23Derived |
| Magistral Small 1.2 24B24B | Q6_K | 20.91 GB | +123.1 GB | 17–23Derived |
| Codestral 22B22B | Q4_K_M | 15.03 GB | +129 GB | 25–33Derived |
| ERNIE 4.5 21B-A3B21B-A3B | Q8_0 | 24.44 GB | +119.6 GB | ≤264Ceiling |
| GPT-OSS 20B21B MoE | MXFP4 | 12.76 GB | +131.2 GB | ≤404Ceiling |
14B · 18
| Model | Quant | Est. VRAM | Headroom | tok/s |
|---|---|---|---|---|
| InternLM2 20B Chat20B | Q5_K_M | 15.59 GB | +128.4 GB | 24–31Derived |
| DeepSeek-Coder-V2-Lite Instruct16B-A2.4B | Q8_0 | 17.56 GB | +126.4 GB | ≤330Ceiling |
| DeepSeek-V2-Lite Chat16B-A2.4B | Q4_K_M | 10.08 GB | +133.9 GB | ≤579Ceiling |
| StarCoder2 15B15B | Q4_K_M | 10.19 GB | +133.8 GB | 37–49Derived |
| Qwen3 14B Instruct14B | Q5_K_M | 11.65 GB | +132.4 GB | 32–42Derived |
| Qwen2.5 14B Instruct14B | Q5_K_M | 11.73 GB | +132.3 GB | 32–43Derived |
| DeepSeek-R1-Distill-Qwen-14B14B | Q4_K_M | 10.14 GB | +133.9 GB | 38–50Derived |
| Phi-4 14B14B | Q5_K_M | 11.77 GB | +132.2 GB | 32–43Derived |
| Phi-3 Medium 14B Instruct14B | Q6_K | 12.86 GB | +131.1 GB | 29–39Derived |
| Mistral Nemo 12B Instruct12B | Q6_K | 11.14 GB | +132.9 GB | 34–44Derived |
| Gemma 3 12B IT12B | Q5_K_M | 9.84 GB | +134.2 GB | 39–51Derived |
| Mellum2.1 12B-A2.5B Thinking12B-A2.5B | Q8_0 | 13.6 GB | +130.4 GB | ≤317Ceiling |
| Stable LM 2 12B Chat12B | Q4_K_M | 8.35 GB | +135.7 GB | 46–61Derived |
| Llama 3.2 11B Vision Instruct11B | Q8_0 | 12.9 GB | +131.1 GB | 29–38Derived |
| Solar 10.7B Instruct11B | Q4_K_M | 7.6 GB | +136.4 GB | 52–69Derived |
| Falcon 3 10B Instruct10B | Q4_K_M | 7.28 GB | +136.7 GB | 53–70Derived |
| Gemma 2 9B Instruct9B | Q8_0 | 11.7 GB | +132.3 GB | 34–45Derived |
| Qwen3.5 9B9B | Q8_0 | 10.37 GB | +133.6 GB | 34–45Derived |
7B · 25
| Model | Quant | Est. VRAM | Headroom | tok/s |
|---|---|---|---|---|
| GLM-4-9B-Chat9B | Q8_0 | 10.16 GB | +133.8 GB | 35–46Derived |
| Qwen3-VL 8B Instruct8B | Q8_0 | 10.39 GB | +133.6 GB | 36–48Derived |
| Granite 4.2 8B8B | Q8_0 | 10.45 GB | +133.6 GB | 36–48Derived |
| Ministral 3 8B Instruct8B | Q8_0 | 10.02 GB | +134 GB | 37–49Derived |
| Qwen2-VL 7B Instruct7B | Q4_K_M | 5.49 GB | +138.5 GB | 67–88Derived |
| Granite 3.1 8B Instruct8B | Q4_K_M | 5.88 GB | +138.1 GB | 68–89Derived |
| Qwen3 8B Instruct8B | Q6_K | 7.64 GB | +136.4 GB | 50–66Derived |
| Llama 3.1 8B Instruct8B | Q8_0 | 9.47 GB | +134.5 GB | 39–52Derived |
| Nous Hermes 3 Llama 3.1 8B8B | Q4_K_M | 5.64 GB | +138.4 GB | 69–91Derived |
| Aya 23 8B8B | Q4_K_M | 5.64 GB | +138.4 GB | 69–91Derived |
| OpenChat 3.6 8B8B | Q4_K_M | 5.64 GB | +138.4 GB | 69–91Derived |
| DeepSeek-R1-Distill-Llama-8B8B | Q5_K_M | 6.51 GB | +137.5 GB | 59–78Derived |
| InternLM2 7B Chat7B | Q4_K_M | 5.45 GB | +138.6 GB | 72–95Derived |
| Qwen2.5 7B Instruct7B | Q6_K | 6.77 GB | +137.2 GB | 54–71Derived |
| Qwen2.5-Coder 7B Instruct7B | Q4_K_M | 5.07 GB | +138.9 GB | 73–96Derived |
| DeepSeek-R1-Distill-Qwen-7B7B | Q4_K_M | 5.07 GB | +138.9 GB | 73–96Derived |
| Gemma 4 E4B ITE4B | Q8_0 | 8.46 GB | +135.5 GB | 42–56Derived |
| OLMo 2 7B Instruct7B | Q8_0 | 10.3 GB | +133.7 GB | 43–57Derived |
| OLMo 3 7B Instruct7B | Q8_0 | 10.3 GB | +133.7 GB | 43–57Derived |
| WizardLM-2 7B7B | Q4_K_M | 5.14 GB | +138.9 GB | 76–101Derived |
| Mistral 7B Instruct v0.37B | Q6_K | 6.75 GB | +137.3 GB | 57–75Derived |
| Zephyr 7B Beta7B | Q6_K | 6.75 GB | +137.3 GB | 57–75Derived |
| Gemma 4 E2B ITE2B | Q8_0 | 5.2 GB | +138.8 GB | ≤171Ceiling |
| Qwen3.5 4B4B | Q8_0 | 4.97 GB | +139 GB | ≤182Ceiling |
| Gemma 3 4B IT4B | Q8_0 | 5.05 GB | +139 GB | ≤184Ceiling |
≤3B · 14
| Model | Quant | Est. VRAM | Headroom | tok/s |
|---|---|---|---|---|
| Qwen3 4B Instruct4B | Q6_K | 4.05 GB | +140 GB | ≤256Ceiling |
| Phi-4 Mini Instruct3.8B | Q8_0 | 4.81 GB | +139.2 GB | ≤206Ceiling |
| Phi-3.5 Mini Instruct3.8B | Q8_0 | 5.89 GB | +138.1 GB | ≤207Ceiling |
| Granite 4.2 3B3B | Q8_0 | 4.41 GB | +139.6 GB | ≤217Ceiling |
| Llama 3.2 3B Instruct3B | Q8_0 | 4.05 GB | +140 GB | ≤247Ceiling |
| Qwen2.5 3B Instruct3B | Q8_0 | 3.59 GB | +140.4 GB | ≤256Ceiling |
| SmolLM3 3B3B | Q8_0 | 3.73 GB | +140.3 GB | ≤257Ceiling |
| Gemma 2 2B Instruct2B | Q8_0 | 3.34 GB | +140.7 GB | ≤304Ceiling |
| Qwen3.5 2B2B | Q8_0 | 2.23 GB | +141.8 GB | ≤404Ceiling |
| Qwen3 1.7B Instruct1.7B | Q8_0 | 2.39 GB | +141.6 GB | ≤460Ceiling |
| Qwen2.5 1.5B Instruct1.5B | Q8_0 | 1.83 GB | +142.2 GB | ≤516Ceiling |
| Llama 3.2 1B Instruct1B | Q8_0 | 1.51 GB | +142.5 GB | ≤640Ceiling |
| Qwen3.5 0.8B0.8B | Q8_0 | 0.94 GB | +143.1 GB | ≤988Ceiling |
| Qwen2.5 0.5B Instruct0.5B | Q8_0 | 0.6 GB | +143.4 GB | ≤1633Ceiling |
How this list is built
Each row is the lowest-perplexity-loss quant of that model whose estimated total — weights plus KV cache at 4K context plus activation buffer — uses at most 88% of the card. That is the calculator's "green" threshold, so every row here has real headroom rather than only just fitting. Raise the context length and the list shortens; the calculator lets you check any combination directly. The speed column is this card: a measured run, a range scaled from this card's own Llama 2 7B benchmark, or a bandwidth ceiling. A figure that only exists for the RTX 4090 is labelled as that card.
Measured on this card
1 benchmark runs in this index were recorded on a Mac M2 Ultra 192G. Everything else on this page is calculated, not observed.
| Model | Stack | tok/s | VRAM used |
|---|---|---|---|
| Llama 3.1 8B | llama.cpp · GGUF Q4_K_M | 90 | 5.7 GB |
Cards with the same budget
What fits is decided by memory, so every 192GB card of this type returns the same list. These pages are not different answers — they differ in throughput, which this index does not measure per card.
Same list, different speed — play one sentence on both cards: Mac M2 Ultra 192G vs Mac M3 Ultra 192G (Qwen3 32B Instruct)
Setup guides for this card
Step-by-step guides written for the Mac M2 Ultra 192G's runtime or memory size, with the commands and the checks that prove the model is running on the card.
Stepping up
A Mac M5 Ultra 512G (512GB) fits 4 more of the indexed models than this card. Mac M5 Ultra 512G →
Common questions
What is the best local LLM for a Mac M2 Ultra 192G?
For everyday use, DBRX Instruct at Q4_K_M — about 84.3 GB of the card's 192 GB at 4K context, leaving 59.7 GB for a longer window. If you want the largest thing that will load, that is Qwen3 235B-A22B Instruct at Q3_K_M (119.6 GB, 24.4 GB spare). "Best" here means best fit for the memory budget — this index does not run task benchmarks, so it cannot tell you which model is smarter.
Can a Mac M2 Ultra 192G run Llama 4 Maverick 17B (128E)?
No. Its smallest build here, Q3_K_M, needs about 203.0 GB and this card has 144.0 GB usable — short by 59.0 GB before any context beyond 4K. The largest model this card does clear is Qwen3 235B-A22B Instruct.
How many tokens per second does a Mac M2 Ultra 192G do on an 8B model at Q4?
This index has 1 measured run on this card — those figures are in the table on this page. What can be stated from specifications: generating a token requires reading every weight once, Llama 3.1 8B Instruct at Q4_K_M is 4.6 GB of weights, and this card moves 800 GB/s — a ceiling near 173 tok/s. Batch size, context length, the runtime and how much of the model sits in cache all take you below it.
Mac M2 Ultra 192G or Mac M3 Ultra 192G for local LLMs?
They hold the same models: 192 GB against 192 GB fits 101 and 101 of 110 respectively at 4K. The difference is throughput — 800 GB/s against 819 GB/s, a 1.0× gap in how fast the weights can be read, which is what token generation is bound by. The Mac M3 Ultra 192G generates faster on any model both can hold.
101 of 110 indexed models fit comfortably in 192GB at 4K context, each at the highest-quality quant that still leaves headroom.
This page's figures change when the model or the runtime does.Last updated 2026-10-18 RSS → /feed.xml