CanItRun Logocanitrun.

Apple M3 Ultra (96GB)

The Apple M3 Ultra (96GB) has 96 GB VRAM and 819 GB/s memory bandwidth. It can run 54 of our 71 tracked models natively in VRAM at 8k context.

With 96 GB LPDDR5X, the Apple M3 Ultra (96GB) is a workstation-tier GPU that can run 54 models natively. It handles 70B-class models at Q4 quantization.

Apple M3 Ultra (96GB): 96GB unified memory at 819 GB/s. Excellent for 70B-class models at high quantization and large MoE inference workloads. One of the fastest Apple Silicon chips for local AI.

Apple M3 Ultra (96GB): 2025 desktop with 96GB unified LPDDR5X at 819 GB/s. 28-core CPU (20P+8E), 60-core GPU, 32-core Neural Engine on TSMC 3nm.

70B at Q4 native ~18-30 t/s, 32B at Q4 ~35-60 t/s. Excellent for 70B-class models at high quantization and large MoE inference workloads.

MLX and llama.cpp Metal fully supported. Thunderbolt 5. One of the fastest Apple Silicon chips for local AI — same bandwidth as the 256GB and 512GB tiers.

VendorApple
ArchitectureApple M3 Ultra
CPU cores28 (20P + 8E)
VRAM96 GB (unified)
Memory typeLPDDR5X
Memory bandwidth819 GB/s
Compute backendMETAL
TierWorkstation
Released2025
Models (native)54 / 71
Models (offload)0 / 71
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

Mixtral 8x22B Instruct v0.1Qwen 3.5 122B-A10B (MoE)Nemotron 3 Super 120BGPT-OSS 120BLlama 4 Scout 109B

Models this GPU runs natively in VRAM (54)

Too large for this GPU (17)

Models mentioned

Mixtral 8x22B Instruct v0.1141B (39B active)
Mistral AIQ4_K_M rec.
Qwen 3.5 122B-A10B (MoE)122B (10B active)
AlibabaQ3_K_M rec.
Nemotron 3 Super 120B120B (12B active)
NVIDIAQ3_K_M rec.
GPT-OSS 120B117B (5B active)
OpenAIQ4_K_M rec.
Llama 4 Scout 109B109B (17B active)
MetaQ4_K_M rec.
GLM-4.5 Air 106B106B (12B active)
Z.aiQ3_K_M rec.
GLM-4.6V 106B106B (12B active)
Z.aiQ3_K_M rec.
Qwen 2.5 72B Instruct72B
AlibabaQ4_K_M rec.
Llama 3.3 70B Instruct70B
MetaQ4_K_M rec.
DeepSeek R1 Distill Llama 70B70B
DeepSeekQ4_K_M rec.
Llama 3.1 70B Instruct70B
MetaQ4_K_M rec.
Mixtral 8x7B Instruct v0.146.7B (12.9B active)
Mistral AIQ4_K_M rec.
Command-R 35B35B
CohereQ4_K_M rec.
Qwen 3.5 35B-A3B (MoE)35B (3B active)
AlibabaQ4_K_M rec.
Qwen 3.6 35B35B
AlibabaQ4_K_M rec.
Yi 1.5 34B Chat34.4B
01.AIQ4_K_M rec.
Qwen3 32B32.8B
AlibabaQ4_K_M rec.
Qwen 2.5 32B Instruct32.5B
AlibabaQ4_K_M rec.
Qwen 2.5 Coder 32B Instruct32.5B
AlibabaQ4_K_M rec.
DeepSeek R1 Distill Qwen 32B32.5B
DeepSeekQ4_K_M rec.
Nemotron 3 Nano 30B32B (3B active)
NVIDIAQ5_K_M rec.
Gemma 4 31B31B
GoogleQ4_K_M rec.
Qwen3 30B-A3B (MoE)30B (3B active)
AlibabaQ4_K_M rec.
Gemma 2 27B Instruct27.2B
GoogleQ4_K_M rec.
Gemma 3 27B Instruct27B
GoogleQ4_K_M rec.
Qwen 3.6 27B27B
AlibabaQ4_K_M rec.
Gemma 4 26B (MoE)26B (3.8B active)
GoogleQ4_K_M rec.
Mistral Small 3.1 24B Instruct24B
Mistral AIQ4_K_M rec.
Mistral Small 22B22.2B
Mistral AIQ4_K_M rec.
GPT-OSS 20B21B (4B active)
OpenAIQ5_K_M rec.
Qwen3 14B14.8B
AlibabaQ5_K_M rec.
Qwen 2.5 14B Instruct14.7B
AlibabaQ5_K_M rec.
Phi-4 14B Instruct14B
MicrosoftQ5_K_M rec.
Mistral Nemo 12B Instruct12.2B
Mistral AIQ5_K_M rec.
Gemma 3 12B Instruct12.2B
GoogleQ5_K_M rec.
Gemma 2 9B Instruct9.2B
GoogleQ5_K_M rec.
Llama 3.1 8B Instruct8B
MetaQ5_K_M rec.
DeepSeek R1 Distill Llama 8B8B
DeepSeekQ5_K_M rec.
Qwen3 8B8B
AlibabaQ5_K_M rec.
Qwen 2.5 7B Instruct7.6B
AlibabaQ6_K rec.
Mistral 7B Instruct v0.37.25B
Mistral AIQ6_K rec.
Gemma 3 4B Instruct4B
GoogleQ6_K rec.
Gemma 4 E4B4B
GoogleQ5_K_M rec.
Phi-3.5 Mini Instruct3.8B
MicrosoftQ6_K rec.
Llama 3.2 3B Instruct3.2B
MetaQ6_K rec.
Qwen 2.5 3B Instruct3.1B
AlibabaQ6_K rec.
Gemma 2 2B Instruct2.6B
GoogleQ8_0 rec.
Gemma 4 E2B2B
GoogleQ8_0 rec.
SmolLM2 1.7B Instruct1.7B
Hugging FaceQ8_0 rec.
Qwen 2.5 1.5B Instruct1.5B
AlibabaQ8_0 rec.
Llama 3.2 1B Instruct1.24B
MetaQ8_0 rec.
Gemma 3 1B Instruct1B
GoogleQ8_0 rec.
Qwen 2.5 0.5B Instruct0.5B
AlibabaQ8_0 rec.
SmolLM2 360M Instruct0.36B
Hugging FaceQ8_0 rec.

Frequently asked questions

How much VRAM does the Apple M3 Ultra (96GB) have?
The Apple M3 Ultra (96GB) has 96 GB of LPDDR5X with 819 GB/s memory bandwidth (unified system memory, shared between CPU and GPU).
What is the Apple M3 Ultra (96GB) best for?
With 96 GB of VRAM, the Apple M3 Ultra (96GB) 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 M3 Ultra (96GB) run locally?
The Apple M3 Ultra (96GB) can run 54 of the 71 open-weight models tracked by CanItRun natively in VRAM at 8k context. Top options include: Llama 3.3 70B Instruct at Q5_K_M, Llama 3.1 8B Instruct at FP32, Llama 3.2 3B Instruct at FP32.
Can the Apple M3 Ultra (96GB) run Llama 3.3 70B Instruct?
Yes. The Apple M3 Ultra (96GB) runs Llama 3.3 70B Instruct natively in VRAM at Q5_K_M quantization, achieving approximately 18.2 tokens per second.
Can the Apple M3 Ultra (96GB) run Qwen 3.6 27B?
Yes. The Apple M3 Ultra (96GB) runs Qwen 3.6 27B natively in VRAM at BF16 quantization, achieving approximately 15.2 tokens per second.
Can the Apple M3 Ultra (96GB) run Llama 3.1 8B Instruct?
Yes. The Apple M3 Ultra (96GB) runs Llama 3.1 8B Instruct natively in VRAM at FP32 quantization, achieving approximately 25.6 tokens per second.