Hardware / Apple /Apple M2 Ultra (76-core GPU)

Apple · Apple Silicon

Apple M2 Ultra (76-core GPU)

for example Mac Studio (2023), Mac Pro (2023). Memory is a build option on these, so not every one of them is sold with 192 GB.

Runs 165 models fully in memory, 16 more with CPU offload. 2 of those sit close enough to 192 GB that our size estimate is what decides them.

Memory

192 GB

Bandwidth

800 GB/s

Runs

165

01

Specifications

Memory

192 GB

Decides which models fit at all.

Memory bandwidth

800 GB/s

How fast it can read the model. Sets the speed of the reply.

Compute

54.4 TFLOPS

Trillions of half-precision sums a second. Sets how fast it reads a long prompt, not how fast it replies.

Power draw

130 W

At full load, for sizing a power supply.

Released

Jun 13, 2023

Architecture

M2

02

What it runs

Best quant per model. Text models first, then speech, then what does not fit.

What is quantisation? →

ModelParamsContextSpeedVerdict

Qwen3.6 35B A3B36B262K262 tok/sFits in memory Gemma 4 26B A4B25.8B262K167 tok/sFits in memory Nemotron 3 Nano 30B A3B31.6B262K262 tok/sFits in memory Ministral 3 3B 25123.8B131K139 tok/sFits in memory gpt-oss-safeguard-20b21.5B131K218 tok/sFits in memory Granite 4.0 Micro3.2B66K289 tok/sFits in memory Qwen3 VL 30B A3B Thinking31.1B262K262 tok/sFits in memory Qwen3 Next 80B A3B Thinking81.3B262K149 tok/sFits in memory gpt-oss-120b120B131K154 tok/sFits in memory Llama 3.2 3B Instruct3.2B131K246 tok/sFits in memory Qwen3.5-35B-A3B36B262K262 tok/sFits in memory Qwen3 VL 30B A3B Instruct31.1B262K223 tok/sFits in memory Qwen3 Next 80B A3B Instruct81.3B262K149 tok/sFits in memory Qwen3 30B A3B Thinking 250730.5B66K149 tok/sFits in memory Qwen3 Coder 30B A3B Instruct30.5B262K223 tok/sFits in memory Qwen3 30B A3B Instruct 250730.5B262K149 tok/sFits in memory Gemma 3 4B4.3B131K215 tok/sFits in memory Nemotron 3.5 Content Safety4.3B131K183 tok/sFits in memory Llama 3.1 8B Instruct8B131K116 tok/sFits in memory UI-TARS 7B8.3B66K111 tok/sFits in memory Qwen3 8B8.2B131K113 tok/sFits in memory Gemma 3n 4B7.8B33K118 tok/sFits in memory Llama 3.2 1B Instruct1.2B33K656 tok/sFits in memory Qwen2.5 7B Instruct7.6B33K122 tok/sFits in memory Granite 4.1 8B8.8B131K105 tok/sFits in memory Ministral 3 8B 25128.9B262K104 tok/sFits in memory Qwen3 VL 8B Thinking8.8B131K105 tok/sFits in memory Qwen3 VL 8B Instruct8.8B262K105 tok/sFits in memory Granite 4.2 8B8.8B131K105 tok/sFits in memory Reka Edge7.1B16K130 tok/sFits in memory Qwen3.5-9B9.7B262K95 tok/sFits in memory Hy-MT2-1.8B2B8K393 tok/sFits in memory Hy-MT2-30B-A3B30.1B8K262 tok/sFits in memory Qwen3.5-122B-A10B125B262K79 tok/sFits in memory Llama 3 8B Lunaris8B8K116 tok/sFits in memory Llama Guard 4 12B12B262K77 tok/sFits in memory Gemma 3 12B12.2B131K76 tok/sFits in memory Rocinante 12B12.2B66K76 tok/sFits in memory Mistral Nemo12.2B131K76 tok/sFits in memory Ministral 3 14B 251213.9B262K67 tok/sFits in memory GLM 4.5 Air111B131K65 tok/sFits in memory Qwen3 14B14.8B131K62 tok/sFits in memory UnslopNemo 12B12.2B33K76 tok/sFits in memory Hy-MT2-7B8B8K98 tok/sFits in memory Hunyuan A13B Instruct80.4B131K60 tok/sFits in memory Reka Flash 320.9B66K44 tok/sFits in memory Cydonia 24B V4.123.6B131K39 tok/sFits in memory Uncensored24B66K39 tok/sFits in memory Mistral Small 3.2 24B24B131K39 tok/sFits in memory Phi 414.7B16K63 tok/sFits in memory Mistral Small 3.1 24B24B66K39 tok/sFits in memory MythoMax 13B13B8K71 tok/sFits in memory Ternary Bonsai 2 27B27B262K34 tok/sFits in memory Qwen3.5-27B27.8B262K33 tok/sFits in memory Gemma 3 27B27.4B262K34 tok/sFits in memory Qwen3.6 27B27.8B262K33 tok/sFits in memory Muse Glimmer 30B29.8B131K31 tok/sFits in memory Qwen3.8 27B27.8B262K33 tok/sFits in memory Gemma 4 31B31.3B66K30 tok/sFits in memory Olmo 3 32B Think32.2B66K29 tok/sFits in memory Qwen3 VL 32B Instruct33.4B131K28 tok/sFits in memory Qwen3 32B32.8B131K28 tok/sFits in memory Mistral Small 323.6B33K39 tok/sFits in memory Nex-N2-Mini35.1B262K22 tok/sFits in memory GLM 4.7 Flash31.2B131K25 tok/sFits in memory ReMM SLERP 13B13B4K71 tok/sFits in memory Laguna XS 2.133.4B262K24 tok/sFits in memory Nemotron 3.5 Lightning31.6B262K25 tok/sFits in memory Nex-N2.5-Mini35.1B262K22 tok/sFits in memory Mixtral 8x22B Instruct141B66K20 tok/sFits in memory Voxtral Small 24B 250724.3B16K38 tok/sFits in memory Qwen2.5 Coder 32B Instruct32.8B33K28 tok/sFits in memory Hermes 4 70B70.6B131K13 tok/sFits in memory Skyfall 36B V236.9B33K25 tok/sFits in memory Qwen2.5 VL 72B Instruct73.4B66K13 tok/sFits in memory Llama 3.3 Euryale 70B70.6B131K11 tok/sFits in memory Llama 3.1 Euryale 70B v2.270.6B131K13 tok/sFits in memory Hermes 3 70B Instruct70.6B131K13 tok/sFits in memory Llama 3.1 70B Instruct70.6B131K13 tok/sFits in memory Llama 3.3 70B Instruct70.6B131K13 tok/sFits in memory Laguna S 2.1118B131K7 tok/sFits in memory Mistral Medium 3.5128B131K7 tok/sFits in memory Ling-2.6-flash107B66K6 tok/sFits in memory Mistral Small 4119B66K7 tok/sFits in memory Nemotron 3 Super124B262K6 tok/sFits in memory Qwen3 Coder Next79.7B262K10 tok/sFits in memory Devstral 2 2512125B131K6 tok/sFits in memory GLM 4.6V108B131K6 tok/sFits in memory Command A111B131K7 tok/sFits in memory GLM 4.5V108B66K6 tok/sFits in memory Llama 4 Scout109B262K9 tok/sFits in memory Ling-3.0-flash128B66K6 tok/sFits in memory Qwen2.5 72B Instruct72.7B33K13 tok/sFits in memory Gemma 2 27B27.2B8K34 tok/sFits in memory Step 3.7 Flash201B33K4 tok/sFits in memoryest Step 3.5 Flash199B33K4 tok/sFits in memoryest Magnum v4 72B72.7B16K11 tok/sFits in memory R1 Distill Llama 70B70.6B8K13 tok/sFits in memory

Speech models it runs67 modelsTranscription and speech models that fit in memory. Throughput figures for a transcription model come from its text decoder, so treat this as a fit answer rather than a speed one.

Niagara 19m Batch.en20M26353 tok/sFits in memory Niagara 38m Batch.en38M13870 tok/sFits in memory Higgs Audio v3 STT2.7B195 tok/sFits in memory Cohere Transcribe 03 20262.1B251 tok/sFits in memory Distil Large v3.50.8B659 tok/sFits in memory Lite Whisper Large v3 Acc1.4B562 tok/sFits in memory Owsm CTC v3.1 1B1.1B840 tok/sFits in memory Owsm CTC v3.2 ft 1B1B527 tok/sFits in memory Owsm CTC v4 1B1B924 tok/sFits in memory Hubert Large Ls960 ft0.3B1647 tok/sFits in memory MMS 1b All1B787 tok/sFits in memory Wav2vec2 Large 960h Lv60 Self0.3B1647 tok/sFits in memory Hojo ASR V15.2B152 tok/sFits in memory Granite 4.0 1b Speech2.3B229 tok/sFits in memory Granite Speech 3.3 2b3B308 tok/sFits in memory Granite Speech 4.1 2b2.3B402 tok/sFits in memory Granite Speech 4.1 2b NAR2.3B402 tok/sFits in memory STT 2.6b en2.6B355 tok/sFits in memory Phi 4 Multimodal Instruct5.6B131K165 tok/sFits in memory Voxtral Mini 3B 25074.7B168 tok/sFits in memory Voxtral Mini 4B Realtime 26024.4B120 tok/sFits in memory Canary 180m Flash0.2B4374 tok/sFits in memory Canary 1b1B527 tok/sFits in memory Canary 1b Flash0.8B659 tok/sFits in memory Canary 1b v21B924 tok/sFits in memory Canary Qwen 2.5b2.6B203 tok/sFits in memory Nemotron 3.5 ASR Streaming 0.6b0.6B1540 tok/sFits in memory Nemotron Speech Streaming en 0.6b0.6B1312 tok/sFits in memory Parakeet CTC 0.6b0.6B1312 tok/sFits in memory Parakeet CTC 1.1b1.1B716 tok/sFits in memory Parakeet RNNT 0.6b0.6B878 tok/sFits in memory Parakeet RNNT 1.1b1.1B716 tok/sFits in memory Parakeet TDT CTC 110m0.1B8397 tok/sFits in memory Parakeet TDT 0.6b v20.6B878 tok/sFits in memory Parakeet TDT 0.6b v30.6B1540 tok/sFits in memory Parakeet TDT 1.1b1.1B479 tok/sFits in memory Nyra Health CrisperWhisper1.6B329 tok/sFits in memory Whisper Large v31.5B525 tok/sFits in memory Whisper Large v3 Turbo0.8B984 tok/sFits in memory MOSS Transcribe Diarize0.9B875 tok/sFits in memory Qwen3 ASR 0.6B0.9B1026 tok/sFits in memory Qwen3 ASR 0.6B HF0.8B1155 tok/sFits in memory Qwen3 ASR 1.7B2.3B342 tok/sFits in memory Qwen3 ASR 1.7B HF2B264 tok/sFits in memory Zipformer cr CTC Transducer XL 290M0.3B2715 tok/sFits in memory Zipformer Transducer XL 290M0.3B1817 tok/sFits in memory ASR Conformer Largescaleasr0.5B1098 tok/sFits in memory GLM ASR Nano 25122.3B402 tok/sFits in memory Granite Speech 5.0 470m Turboctc0.5B1575 tok/sFits in memory Granite Speech 5.0 470m Turboctc nc0.5B1847 tok/sFits in memory Fun ASR Nano 2512 HF0.8B1155 tok/sFits in memory Gemma 4 E2B it5.1B181 tok/sFits in memory Niagara 9m Batch.en9M58562 tok/sFits in memory Niagara 84m Batch.en84M10997 tok/sFits in memory Gemma 4 E4B it8B131K116 tok/sFits in memory VibeVoice ASR HF8.3B111 tok/sFits in memory ARK ASR 0.6B1.3B33K606 tok/sFits in memory ARK ASR 3B4.1B33K225 tok/sFits in memory Audio8 ASR 0.1B0.3B33K1757 tok/sFits in memory Granite Speech 3.3 8b8.6B107 tok/sFits in memory MOSS Transcribe Preview 2B2.4B33K385 tok/sFits in memory Higgs Audio v3 8b STT v28.9B104 tok/sFits in memory Gemma 4 12B it12B77 tok/sFits in memory Moonshine Streaming Medium0.3B4K1757 tok/sFits in memory Moonshine Streaming Small0.1B4K7874 tok/sFits in memory Moonshine Streaming Tiny0B4Kno speed estimateFits in memory Moonshine Tiny0B194no speed estimateFits in memory

Runs only by spilling into system memory16 modelsThese load, but part of the weights sits in ordinary system memory, which is far slower than the chip.

Hy3299B—no speed estimateSpills to system RAM MiniMax M2.7229B—no speed estimateSpills to system RAM Qwen3 VL 235B A22B Thinking236B—no speed estimateSpills to system RAM Laguna M.1226B—no speed estimateSpills to system RAM DeepSeek V4 Flash291B—no speed estimateSpills to system RAM Hy3 preview299B—no speed estimateSpills to system RAM MiMo-V2.5311B—no speed estimateSpills to system RAMest MiniMax M2.5229B—no speed estimateSpills to system RAM MiniMax M2.1229B—no speed estimateSpills to system RAMest Qwen3 VL 235B A22B Instruct236B—no speed estimateSpills to system RAMest Qwen3 235B A22B Thinking 2507235B—no speed estimateSpills to system RAM Qwen3 235B A22B Instruct 2507235B—no speed estimateSpills to system RAM MiniMax M2229B—no speed estimateSpills to system RAMest Inkling Small266B—no speed estimateSpills to system RAM DeepSeek V4 Flash Vision Exp305B—no speed estimateSpills to system RAMest GLM 5.3 Flash321B—no speed estimateSpills to system RAMest

Too large for this device43 modelsThe weights do not fit even with part of them offloaded to system memory. Closest calls first.

Nex-N2-Pro397B—no speed estimateToo large Nex-N2.5-Pro397B—no speed estimateToo large Trinity Large Thinking399B—no speed estimateToo large Hermes 4 405B406B—no speed estimateToo large Hermes 3 405B Instruct406B—no speed estimateToo large GLM 4.6357B—no speed estimateToo large GLM 4.7358B—no speed estimateToo large GLM 4.5358B—no speed estimateToo large Llama 4 Maverick402B—no speed estimateToo large Qwen3 Coder 480B A35B480B—no speed estimateToo large ERNIE 4.5 VL 424B A47B424B—no speed estimateToo large MiniMax M3427B—no speed estimateToo large MiniMax-01456B—no speed estimateToo large Nemotron 3 Ultra561B—no speed estimateToo large Jamba Large 1.7399B—no speed estimateToo large Qwen3.5 397B A17B403B—no speed estimateToo large GLM 5.1754B—no speed estimateToo large GLM 5.2753B—no speed estimateToo large MiniMax M1456B—no speed estimateToo large DeepSeek V3.1 Terminus685B—no speed estimateToo large DeepSeek V3685B—no speed estimateToo large DeepSeek V3.1685B—no speed estimateToo large DeepSeek V4.1 Flash763B—no speed estimateToo large MiMo-V2.5-Pro1T—no speed estimateToo large Kimi K2 07111T—no speed estimateToo large Kimi K2.51.1T—no speed estimateToo large R1 0528685B—no speed estimateToo large DeepSeek V3 0324685B—no speed estimateToo large DeepSeek V3.2685B—no speed estimateToo large DeepSeek V3.2 Exp685B—no speed estimateToo large Kimi K2 09051T—no speed estimateToo large Ring-2.6-1T1T—no speed estimateToo large Ling-2.6-1T1T—no speed estimateToo large Kimi K2.7 Code1.1T—no speed estimateToo large Kimi K2.61.1T—no speed estimateToo large GLM 5754B—no speed estimateToo large Hy4 preview780B—no speed estimateToo large Inkling952B—no speed estimateToo large Kimi K2 Thinking1.1T—no speed estimateToo large LongCat 2.01.8T—no speed estimateToo large DeepSeek V4 Pro1.6T—no speed estimateToo large Qwen3.8 2.4T A95B2.4T—no speed estimateToo large Kimi K32.8T—no speed estimateToo large

224 models have a verdict on this device: 165 fit in memory, 16 spill into system memory and 43 are too large. Every one of them is on this page.

03

Other M2 chips

The same generation in other memory sizes. Memory is what decides which models fit.

Apple M2 (8-core GPU, 8GB unified)8 GB · less memory Apple M2 (10-core GPU)24 GB · less memory Apple M2 Pro (19-core GPU)32 GB · less memory Apple M2 Max (38-core GPU)96 GB · less memory

Something wrong on this page? Tell us