作者： Jakub Rusinowski · 最后更新： 2026年7月12日
"Strix Halo" SoC pairing 16 Zen 5 CPU cores with a 40-CU RDNA 3.5 iGPU (Radeon 8060S), sharing up to 128 GB of LPDDR5X-8000 memory at 256 GB/s. Up to 96 GB can be allocated to the GPU as VRAM via AMD Variable Graphics Memory (Windows) or GTT (Linux) — the figure used here. Powers AI mini-PCs like the Framework Desktop, GMKtec EVO-X2, and HP Z2 Mini G1a. ROCm support is improving but still behind CUDA.
| VRAM | 96 GB |
| Memory Bandwidth | 256 GB/s |
| TDP | 120 W |
| Architecture | Zen 5 + RDNA 3.5 "Strix Halo" |
| Release Year | 2025 |
| MSRP at Launch | $1,999 |
| Inference Speed (Llama 3.1 8B Q4_K_M) | 18–38 tok/s (estimated) |
| Inference Speed (Llama 3.3 70B Q4_K_M) | 2.4–5.1 tok/s (estimated) |
As an Amazon Associate we earn from qualifying purchases. Cloud GPU links are referral links — we may earn a commission at no extra cost to you.
All models below run comfortably in 96 GB VRAM with Q4_K_M quantization.
| Llama 3.3 | Llama 3.3 70B Instruct · 43 GB VRAM · Q4_K_M · ollama run llama3.3 |
| Llama 3.1 Family | Llama 3.1 8B Instruct · 6 GB VRAM · Q4_K_M · ollama run llama3.1 |
| Llama 4 | Llama 4 Scout 17B · 67 GB VRAM · Q4_K_M · ollama run llama4:scout |
| DeepSeek R1 | DeepSeek R1 Distill Qwen 32B · 20 GB VRAM · Q4_K_M · ollama run deepseek-r1:32b |
| Qwen 3 | Qwen 3 32B · 21 GB VRAM · Q4_K_M · ollama run qwen3:32b |
| Qwen3-Coder | Qwen3-Coder 80B-A3B (MoE) · 49 GB VRAM · Q4_K_M · ollama run qwen3-coder:80b-a3b-q4 |
| Gemma 3 | Gemma 3 27B Instruct · 17 GB VRAM · Q4_K_M · ollama run gemma3:27b |
| Mistral Small 3.1 | Mistral Small 3.1 24B · 15 GB VRAM · Q4_K_M · ollama run mistral-small3.1 |
Install Ollama then run the recommended model for this GPU:
ollama run llama3.3:70b
Yes — the AMD Ryzen AI Max+ 395 has 96 GB VRAM and runs "Strix Halo" SoC pairing 16 Zen 5 CPU cores with a 40-CU RDNA 3.5 iGPU (Radeon 8060S), sharing up to 128 GB of LPDDR5X-8
The AMD Ryzen AI Max+ 395 is estimated to run Llama 3.1 8B at 18–38 tok/s with Q4_K_M quantization. For Llama 3.3 70B the estimate is 2.4–5.1 tok/s. These are modelled estimates, not measurements — see /en/methodology.
With 96 GB you can run: Llama 3.3, Llama 3.1 Family, Llama 4, DeepSeek R1, Qwen 3. Use Ollama for the easiest setup: ollama run llama3.3:70b.
← All GPU Reviews | All Hardware | Check Your Hardware | Full Benchmarks | Can I Run It?