Device profile · Windows
Best local LLMs for AMD Ryzen AI Halo (128GB)
AMD Ryzen AI Halo (128GB) has ~96 GB usable for model weights and runs 118 of 135 popular models. Best tool: LM Studio.
117 run well 1 tight fit 17 too heavy
- Usable memory
- ~96 GB
- Models run
- 118
- Too large
- 17
- Top pick
- 30.5B
Top pick Q4_K_M
Qwen3 30B-A3B Yes
Runs at Q4_K_M using ~20.7 GB of ~96 GB usable. You have room for FP16 for higher quality.
Runs on AMD Ryzen AI Halo (128GB)
Compatible models 118 total
- NE YesNemotron 3 Super 120B-A12B120B MoE · ~80.3 GB at Q4_K_M
- LA TightLaguna S 2.1118B MoE · ~92.8 GB at Q4_K_M
- Yesgpt-oss 120B117B MoE · ~62.4 GB at Q4_K_M
- CR YesCommand A111B · ~70.4 GB at Q4_K_M
- YesLlama 4 Scout109B MoE · ~64.2 GB at Q4_K_M
- GL YesGLM-4.5-Air106B MoE · ~71.8 GB at Q4_K_M
- S YesSarvam-105B105B MoE · ~67.5 GB at Q4_K_M
- HY YesHunyuan-A13B-Instruct80B MoE · ~48.3 GB at Q4_K_M
- YesQwen2.5 72B72B · ~50.2 GB at Q4_K_M
- YesLlama 3.3 70B70B · ~45.3 GB at Q4_K_M
- NE YesLlama-3.3-Nemotron-Super-49B-v149B · ~30.6 GB at Q4_K_M
- YesMixtral 8x7B46.7B MoE · ~28.9 GB at Q4_K_M
- SE YesSeed-OSS 36B Instruct36B · ~22.5 GB at Q4_K_M
- YesQwen3.6 35B-A3B36B MoE · ~24.5 GB at Q4_K_M
- CR YesCommand R 35B35B · ~22.3 GB at Q4_K_M
- OR YesOrnith 1.0 35B35B MoE · ~23.2 GB at Q4_K_M
- YesQwen-AgentWorld 35B-A3B34.7B MoE · ~24.3 GB at Q4_K_M
- Yi YesYi 1.5 34B34B · ~21.4 GB at Q4_K_M
- FN YesFalcon-H1-34B-Instruct34B · ~21.1 GB at Q4_K_M
- LA YesLaguna XS 2.133.4B MoE · ~22.2 GB at Q4_K_M
- YesGemma 4 31B32.7B · ~20.5 GB at Q4_K_M
- YesQwen2.5 32B32B · ~22.1 GB at Q4_K_M
- YesQwen3 32B32B · ~22 GB at Q4_K_M
- YesDeepSeek-R1-Distill-Qwen 32B32B · ~22.1 GB at Q4_K_M
- YesQwen2.5 Coder 32B32B · ~20.7 GB at Q4_K_M
- YesGranite 4.0 H Small32B MoE · ~20.4 GB at Q4_K_M
- GL YesGLM-4-32B-041432B · ~20.5 GB at Q4_K_M
- EX YesEXAONE 4.0 32B32B · ~20.2 GB at Q4_K_M
- OL YesOLMo 2 32B Instruct32B · ~21.7 GB at Q4_K_M
- YesGranite 4.0 H Small32B MoE · ~20.3 GB at Q4_K_M
- OL YesOlmo 3.1 32B Instruct32B · ~20.3 GB at Q4_K_M
- YesQwen3 30B-A3B30.5B MoE · ~20.7 GB at Q4_K_M
- YesQwen3-Coder 30B-A3B30.5B MoE · ~19.4 GB at Q4_K_M
- NO YesNorth Mini Code 1.030.5B MoE · ~21.1 GB at Q4_K_M
- S YesSarvam-30B30B MoE · ~21.7 GB at Q4_K_M
- NE YesNemotron 3 Nano 30B-A3B30B MoE · ~25.1 GB at Q4_K_M
- NE YesNemotron Cascade 2 30B-A3B30B MoE · ~25.1 GB at Q4_K_M
- GL YesGLM-4.7-Flash30B MoE · ~19.2 GB at Q4_K_M
- YesGranite 4.1 30B28.9B · ~19.1 GB at Q4_K_M
- YesQwen3.6 27B27.8B · ~18.9 GB at Q4_K_M
- YesGemma 2 27B27B · ~18.7 GB at Q4_K_M
- YesGemma 3 27B27B · ~18.6 GB at Q4_K_M
- YesGemma 4 26B-A4B26.5B MoE · ~19 GB at Q4_K_M
- YesMistral Small 3 24B24B · ~16.3 GB at Q4_K_M
- S YesSarvam-M 24B24B · ~16.3 GB at Q4_K_M
- YesMistral Small 3.1 24B24B · ~15.4 GB at Q4_K_M
- YesMagistral Small24B · ~16 GB at Q4_K_M
- YesDevstral Small24B · ~15.4 GB at Q4_K_M
- LF YesLFM2 24B-A2B24B MoE · ~15.4 GB at Q4_K_M
- Yesgpt-oss 20B21B MoE · ~13.2 GB at Q4_K_M
- ER YesERNIE 4.5 21B-A3B21B MoE · ~14.3 GB at Q4_K_M
- YesDeepSeek-V2-Lite16B MoE · ~12.2 GB at Q4_K_M
- YesPhi-4 14B14B · ~10.8 GB at Q4_K_M
- YesQwen2.5 14B14B · ~10.7 GB at Q4_K_M
- YesQwen3 14B14B · ~10.7 GB at Q4_K_M
- YesDeepSeek-R1-Distill-Qwen 14B14B · ~10.7 GB at Q4_K_M
- YesQwen2.5 Coder 14B14B · ~10.1 GB at Q4_K_M
- YesPhi-4-reasoning14B · ~10.1 GB at Q4_K_M
- YesMistral Nemo 12B12.2B · ~8.6 GB at Q4_K_M
- YesGemma 3 12B12B · ~8.9 GB at Q4_K_M
- YesGemma 4 12B12B · ~8.7 GB at Q4_K_M
- YesLlama 3.2 Vision 11B10.7B · ~9 GB at Q4_K_M
- FN YesFalcon3 10B10B · ~7.5 GB at Q4_K_M
- YesGemma 2 9B9B · ~7.3 GB at Q4_K_M
- GL YesGLM-4 9B9B · ~7.3 GB at Q4_K_M
- GL YesGLM-4-9B-04149B · ~7.2 GB at Q4_K_M
- NE YesNemotron Nano 9B v29B · ~7.6 GB at Q4_K_M
- OR YesOrnith 1.0 9B9B · ~7.1 GB at Q4_K_M
- YesGranite 4.1 8B8.8B · ~6.8 GB at Q4_K_M
- LF YesLFM2.5 8B-A1B8.3B MoE · ~6.7 GB at Q4_K_M
- YesQwen2.5-VL 7B8.29B · ~7.1 GB at Q4_K_M
- YesDeepSeek-R1-0528-Qwen3-8B8.19B · ~6.2 GB at Q4_K_M
- YesLlama 3.1 8B8B · ~6.4 GB at Q4_K_M
- YesQwen3 8B8B · ~6.5 GB at Q4_K_M
- YesDeepSeek-R1-Distill-Llama 8B8B · ~6.4 GB at Q4_K_M
- YesGemma 3n E4B8B · ~5.7 GB at Q4_K_M
- YesGemma 4 E4B8B · ~6.5 GB at Q4_K_M
- YesMistral 7B7B · ~5.8 GB at Q4_K_M
- YesQwen2.5 7B7B · ~6.1 GB at Q4_K_M
- YesDeepSeek-R1-Distill-Qwen 7B7B · ~6.1 GB at Q4_K_M
- YesQwen2.5 Coder 7B7B · ~5.8 GB at Q4_K_M
- YesGemma 4 E2B5.1B · ~4.4 GB at Q4_K_M
- YesGemma 3 4B4B · ~3.8 GB at Q4_K_M
- YesQwen3 4B4B · ~3.8 GB at Q4_K_M
- NE YesNemotron 3 Nano 4B4B · ~3.9 GB at Q4_K_M
- YesPhi-3.5-mini 3.8B3.82B · ~3.7 GB at Q4_K_M
- YesPhi-4-mini 3.8B3.8B · ~3.8 GB at Q4_K_M
- YesPhi-4-mini-reasoning3.8B · ~3.6 GB at Q4_K_M
- YesQwen2.5-VL 3B3.75B · ~4.4 GB at Q4_K_M
- YesGranite 4.1 3B3.4B · ~3.3 GB at Q4_K_M
- YesQwen2.5 3B3.09B · ~3.3 GB at Q4_K_M
- YesQwen2.5 Coder 3B3.09B · ~3 GB at Q4_K_M
- YesLlama 3.2 3B3B · ~3.2 GB at Q4_K_M
- YesSmolLM3 3B3B · ~3 GB at Q4_K_M
- OP YesApple OpenELM 3B3B · ~3 GB at Q4_K_M
- YesGemma 2 2B2.61B · ~2.9 GB at Q4_K_M
- YesGranite 3.1 2B2.53B · ~2.8 GB at Q4_K_M
- S YesSarvam-1 2B2B · ~2.7 GB at Q4_K_M
- YesSmolLM2 1.7B1.7B · ~2.2 GB at Q4_K_M
- YesQwen3 1.7B1.7B · ~2.4 GB at Q4_K_M
- YesQwen2.5 1.5B1.54B · ~2.2 GB at Q4_K_M
- YesQwen2.5 Coder 1.5B1.54B · ~2 GB at Q4_K_M
- MI YesMiniCPM-V 4.61.3B · ~1.6 GB at Q4_K_M
- LF YesLFM2 1.2B1.17B · ~2.5 GB at Q4_K_M
- LF YesLFM2.5 1.2B1.17B · ~1.8 GB at Q4_K_M
- LF YesLFM2.5 1.2B Thinking1.17B · ~1.8 GB at Q4_K_M
- TL YesTinyLlama 1.1B1.1B · ~1.8 GB at Q4_K_M
- OP YesApple OpenELM 1.1B1.1B · ~1.7 GB at Q4_K_M
- YesLlama 3.2 1B1B · ~1.8 GB at Q4_K_M
- YesGemma 3 1B1B · ~1.8 GB at Q4_K_M
- LF YesLFM2 700M0.742B · ~1.9 GB at Q4_K_M
- YesQwen3 0.6B0.6B · ~1.5 GB at Q4_K_M
- YesQwen2.5 0.5B0.494B · ~1.5 GB at Q4_K_M
- YesQwen2.5 Coder 0.5B0.494B · ~1.4 GB at Q4_K_M
- YesSmolLM2 360M0.362B · ~1.2 GB at Q4_K_M
- LF YesLFM2 350M0.354B · ~1.4 GB at Q4_K_M
- YesGemma 3 270M0.27B · ~1.1 GB at Q4_K_M
- YesSmolLM2 135M0.135B · ~1 GB at Q4_K_M
Too large for this device
Best way to run models on Windows
Runtime guide Windows
Beginner: LM Studio, Best GUI on Windows, auto-detects CUDA/Vulkan backends.
Power user: Ollama (CUDA), Scriptable server; CUDA path is fastest on NVIDIA.
AMD GPUs run via Vulkan/ROCm at roughly half CUDA throughput. NVIDIA is the smooth path on Windows.
Full Windows tool guide →FAQ
What is the best local LLM for AMD Ryzen AI Halo (128GB)?
Qwen3 30B-A3B is the strongest model that runs comfortably, using ~20.7 GB at Q4_K_M of the ~96 GB usable on AMD Ryzen AI Halo (128GB).
How much of AMD Ryzen AI Halo (128GB)'s memory can I use for a model?
About 96 GB. On a CPU-only machine, leave headroom for the OS and apps.
Which tool should I use on Windows?
LM Studio (Best GUI on Windows, auto-detects CUDA/Vulkan backends.) or Ollama (CUDA) for speed. AMD GPUs run via Vulkan/ROCm at roughly half CUDA throughput. NVIDIA is the smooth path on Windows.
Compare nearby devices
Hardware with a similar usable pool, and how many of the 135 tracked models each runs next to AMD Ryzen AI Halo (128GB)'s 118.
Apple M4 Max (128GB)
~96 GB usable · runs 118 of 135
Apple M5 Max (128GB)
~96 GB usable · runs 118 of 135
Apple M4 Max (64GB)
~48 GB usable · runs 109 of 135
Apple M4 Pro (48GB)
~32 GB usable · runs 108 of 135
Apple M5 Pro (48GB)
~32 GB usable · runs 108 of 135
Nvidia GeForce RTX 5090 (32GB)
~31 GB usable · runs 108 of 135
Sources
Memory figures are estimates. See methodology.