Apple M2 Ultra (384GB)
The Apple M2 Ultra (384GB) has 384 GB VRAM and 800 GB/s memory bandwidth. It can run 69 of our 70 tracked models natively in VRAM at 8k context.
With 384 GB LPDDR5, the Apple M2 Ultra (384GB) is a workstation-tier GPU that can run 69 models natively. It handles 70B-class models at Q4 quantization.
The Apple M2 Ultra (384GB) is the maximum-memory configuration of Apple's M2 Ultra chip. Its 384GB unified memory pool at 800 GB/s is sufficient to run 405B parameter models at moderate quantization without any CPU offloading — a capability previously exclusive to multi-GPU server racks. As a 2023 chip, it has since been superseded by the M4 Ultra, but remains a powerful and fully-supported platform for local large-model inference with MLX and llama.cpp.
Apple M2 Ultra (384GB): 2023 workstation with 384GB LPDDR5 at 800 GB/s. 24-core CPU (16P+8E), 76-core GPU.
405B at Q2-Q3 native. 70B at Q4 ~13-16 t/s decode. Most extensively benchmarked Apple chip for LLM.
Excellent llama.cpp K-quants support. Block-based K-quants (Q2_K, Q4_K) outperform IQ-quants by 30-45%. $0.376/hour vs $0.665/hour 2xRTX A6000.
| Vendor | Apple |
| Architecture | Apple M2 Ultra |
| CPU cores | 24 (16P + 8E) |
| VRAM | 384 GB (unified) |
| Memory type | LPDDR5 |
| Memory bandwidth | 800 GB/s |
| Compute backend | METAL |
| Tier | Workstation |
| Released | 2023 |
| Models (native) | 69 / 70 |
| Models (offload) | 0 / 70 |
Software: MLX gives the best performance on Apple Silicon; llama.cpp Metal backend is a solid alternative. Both are well-supported by Ollama.
Popular models for this GPU
Models this GPU runs natively in VRAM (69)
- Kimi K2.61000B · MMLU-Pro 87.2Q2_K · ~83.6 t/s
- GLM-5.1 754B754B · MMLU-Pro 86.5Q3_K_M · ~46.5 t/s
- GLM-5 744B744B · MMLU-Pro 85.7Q3_K_M · ~51.2 t/s
- DeepSeek V3 671B671B · MMLU-Pro 75.9Q3_K_M · ~55.3 t/s
- DeepSeek R1 671B671B · MMLU-Pro 85.0Q3_K_M · ~55.3 t/s
- MiniMax M1 456B456B · MMLU-Pro 81.1Q5_K_M · ~29.7 t/s
- Llama 3.1 405B Instruct405B · MMLU-Pro 73.3Q6_K · ~2.4 t/s
- Llama 4 Maverick 400B400B · MMLU-Pro 80.5Q6_K · ~63.1 t/s
- GLM-4.7 358B358B · MMLU-Pro 84.3Q6_K · ~33.5 t/s
- GLM-4.5 355B355B · MMLU-Pro 84.6Q6_K · ~33.5 t/s
- GLM-4.6 355B355B · MMLU-Pro 84.5Q6_K · ~33.5 t/s
- DeepSeek V4 Flash 284B284B · MMLU-Pro 86.3Q8_0 · ~67.7 t/s
- Qwen3 235B-A22B (MoE)235B · MMLU-Pro 84.4Q8_0 · ~40 t/s
- MiniMax M2.5 229B229B · MMLU-Pro 84.8Q8_0 · ~88 t/s
- MiniMax M2.7 229B229B · MMLU-Pro 86.0Q8_0 · ~88 t/s
- Mixtral 8x22B Instruct v0.1141B · MMLU-Pro 40.0BF16 · ~11.3 t/s
- Qwen 3.5 122B-A10B (MoE)122B · MMLU-Pro 86.7BF16 · ~44 t/s
- Nemotron 3 Super 120B120B · MMLU-Pro 83.7BF16 · ~36.7 t/s
- GPT-OSS 120B117B · MMLU-Pro 80.7BF16 · ~88 t/s
- Llama 4 Scout 109B109B · MMLU-Pro 74.3BF16 · ~25.9 t/s
- GLM-4.5 Air 106B106B · MMLU-Pro 81.4BF16 · ~36.7 t/s
- GLM-4.6V 106B106B · MMLU-Pro 79.9BF16 · ~36.7 t/s
- Qwen 2.5 72B Instruct72B · MMLU-Pro 71.1FP32 · ~2.8 t/s
- Llama 3.3 70B Instruct70B · MMLU-Pro 68.9FP32 · ~2.9 t/s
- DeepSeek R1 Distill Llama 70B70B · MMLU-Pro 70.0FP32 · ~2.9 t/s
- Llama 3.1 70B Instruct70B · MMLU-Pro 66.4FP32 · ~2.9 t/s
- Mixtral 8x7B Instruct v0.146.7B · MMLU-Pro 29.7FP32 · ~17.1 t/s
- Command-R 35B35B · MMLU-Pro 33.0FP32 · ~5.7 t/s
- Qwen 3.5 35B-A3B (MoE)35B · MMLU-Pro 84.2FP32 · ~73.3 t/s
- Qwen 3.6 35B35B · MMLU-Pro 85.2FP32 · ~5.7 t/s
- Yi 1.5 34B Chat34.4B · MMLU-Pro 37.0FP32 · ~5.8 t/s
- Qwen3 32B32.8B · MMLU-Pro 65.5FP32 · ~6.1 t/s
- Qwen 2.5 32B Instruct32.5B · MMLU-Pro 69.0FP32 · ~6.2 t/s
- Qwen 2.5 Coder 32B Instruct32.5B · MMLU-Pro 50.4FP32 · ~6.2 t/s
- DeepSeek R1 Distill Qwen 32B32.5B · MMLU-Pro 65.0FP32 · ~6.2 t/s
- Nemotron 3 Nano 30B32B · MMLU-Pro 78.3FP32 · ~73.3 t/s
- Gemma 4 31B31B · MMLU-Pro 85.2FP32 · ~6.5 t/s
- Qwen3 30B-A3B (MoE)30B · MMLU-Pro 61.5FP32 · ~73.3 t/s
- Gemma 2 27B Instruct27.2B · MMLU-Pro 38.0FP32 · ~7.4 t/s
- Gemma 3 27B Instruct27B · MMLU-Pro 67.5FP32 · ~7.4 t/s
- Qwen 3.6 27B27B · MMLU-Pro 86.2FP32 · ~7.4 t/s
- Gemma 4 26B (MoE)26B · MMLU-Pro 82.6FP32 · ~57.9 t/s
- Mistral Small 3.1 24B Instruct24B · MMLU-Pro 66.8FP32 · ~8.3 t/s
- Mistral Small 22B22.2B · MMLU-Pro 49.2FP32 · ~9 t/s
- GPT-OSS 20B21B · MMLU-Pro 67.9FP32 · ~55 t/s
- Qwen3 14B14.8B · MMLU-Pro 61.0FP32 · ~13.5 t/s
- Qwen 2.5 14B Instruct14.7B · MMLU-Pro 63.7FP32 · ~13.6 t/s
- Phi-4 14B Instruct14B · MMLU-Pro 70.4FP32 · ~14.3 t/s
- Mistral Nemo 12B Instruct12.2B · MMLU-Pro 35.6FP32 · ~16.4 t/s
- Gemma 3 12B Instruct12.2B · MMLU-Pro 60.6FP32 · ~16.4 t/s
- Gemma 2 9B Instruct9.2B · MMLU-Pro 32.0FP32 · ~21.7 t/s
- Llama 3.1 8B Instruct8B · MMLU-Pro 48.3FP32 · ~25 t/s
- DeepSeek R1 Distill Llama 8B8B · MMLU-Pro 41.0FP32 · ~25 t/s
- Qwen3 8B8B · MMLU-Pro 56.7FP32 · ~25 t/s
- Qwen 2.5 7B Instruct7.6B · MMLU-Pro 56.3FP32 · ~26.3 t/s
- Mistral 7B Instruct v0.37.25B · MMLU-Pro 30.0FP32 · ~27.6 t/s
- Gemma 3 4B Instruct4B · MMLU-Pro 43.6FP32 · ~50 t/s
- Gemma 4 E4B4B · MMLU-Pro 69.4FP32 · ~50 t/s
- Phi-3.5 Mini Instruct3.8B · MMLU-Pro 47.4FP32 · ~52.6 t/s
- Llama 3.2 3B Instruct3.2B · MMLU-Pro 24.0FP32 · ~62.5 t/s
- Qwen 2.5 3B Instruct3.1B · MMLU-Pro 32.4FP32 · ~64.5 t/s
- Gemma 2 2B Instruct2.6B · MMLU-Pro 17.8FP32 · ~76.9 t/s
- Gemma 4 E2B2B · MMLU-Pro 60.0FP32 · ~100 t/s
- SmolLM2 1.7B Instruct1.7B · MMLU-Pro 19.0FP32 · ~117.6 t/s
- Qwen 2.5 1.5B Instruct1.5B · MMLU-Pro 16.8FP32 · ~133.3 t/s
- Llama 3.2 1B Instruct1.24B · MMLU-Pro 12.5FP32 · ~161.3 t/s
- Gemma 3 1B Instruct1B · MMLU-Pro 14.7FP32 · ~200 t/s
- Qwen 2.5 0.5B Instruct0.5B · MMLU-Pro 10.0FP32 · ~400 t/s
- SmolLM2 360M Instruct0.36B · MMLU-Pro 8.0FP32 · ~555.6 t/s
Too large for this GPU (1)
Frequently asked questions
- How much VRAM does the Apple M2 Ultra (384GB) have?
- The Apple M2 Ultra (384GB) has 384 GB of LPDDR5 with 800 GB/s memory bandwidth (unified system memory, shared between CPU and GPU).
- What is the Apple M2 Ultra (384GB) best for?
- With 384 GB of VRAM, the Apple M2 Ultra (384GB) is a server-class GPU designed for running the largest open-weight models (70B–405B) at high quantization with ample context.
- What LLMs can the Apple M2 Ultra (384GB) run locally?
- The Apple M2 Ultra (384GB) can run 69 of the 70 open-weight models tracked by CanItRun natively in VRAM at 8k context. Top options include: Llama 3.3 70B Instruct at FP32, Llama 3.1 8B Instruct at FP32, Llama 3.2 3B Instruct at FP32.
- Can the Apple M2 Ultra (384GB) run Llama 3.3 70B Instruct?
- Yes. The Apple M2 Ultra (384GB) runs Llama 3.3 70B Instruct natively in VRAM at FP32 quantization, achieving approximately 2.9 tokens per second.
- Can the Apple M2 Ultra (384GB) run Qwen 3.6 27B?
- Yes. The Apple M2 Ultra (384GB) runs Qwen 3.6 27B natively in VRAM at FP32 quantization, achieving approximately 7.4 tokens per second.
- Can the Apple M2 Ultra (384GB) run Llama 3.1 8B Instruct?
- Yes. The Apple M2 Ultra (384GB) runs Llama 3.1 8B Instruct natively in VRAM at FP32 quantization, achieving approximately 25 tokens per second.