Best local LLMs for AMD Ryzen AI Max+ PRO 495 (Ryzen AI Halo) 192GB
AMD Ryzen AI Max+ PRO 495 (Ryzen AI Halo) 192GB. Memory-fit math with KV cache, speeds labelled by how they were obtained, and commands you can paste.
Best for Coding · 16k context
| Model | Quant | Fit | Memory | Speed (generation) |
|---|---|---|---|---|
| Nemotron 3.5 Super VL 120B-A12B | Q4_K_M | fits fully | 70.4 GB | ~25 tok/s estimated |
| Qwen3.8 Flash Next | UD-Q5_K_XL | fits fully | 158.8 GB | ~33 tok/s estimated |
| gpt-oss 120B | Q8_0 | fits fully | 64.0 GB | ~62 tok/s estimated |
| Qwen3-Coder 30B-A3B | Q8_0 | fits fully | 34.1 GB | ~49 tok/s estimated |
| Laguna XS 2.1 | Q8_0 | fits fully | 36.3 GB | ~54 tok/s estimated |
llama-server -hf lmstudio-community/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-GGUF:Q4_K_M -c 16384 -fa on --jinja -ngl 99
OLLAMA_CONTEXT_LENGTH=16384 ollama serve ollama run hf.co/lmstudio-community/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-GGUF:Q4_K_M
Context length is an Ollama server setting: quit the Ollama app or stop the service first, or set it in its environment (launchctl on Mac, systemctl on Linux).
Best for Chat / general · 16k context
| Model | Quant | Fit | Memory | Speed (generation) |
|---|---|---|---|---|
| Nemotron 3.5 Super VL 120B-A12B | Q4_K_M | fits fully | 70.4 GB | ~25 tok/s estimated |
| Qwen3.8 Flash Next | UD-Q5_K_XL | fits fully | 158.8 GB | ~33 tok/s estimated |
| gpt-oss 120B | Q8_0 | fits fully | 64.0 GB | ~62 tok/s estimated |
| Gemma 4 12B | Q8_0 | fits fully | 13.3 GB | ~17 tok/s estimated |
| Nemotron 3 Nano 30B-A3B | Q8_0 | fits fully | 33.7 GB | ~46 tok/s estimated |
llama-server -hf lmstudio-community/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-GGUF:Q4_K_M -c 16384 -fa on --jinja -ngl 99
OLLAMA_CONTEXT_LENGTH=16384 ollama serve ollama run hf.co/lmstudio-community/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-GGUF:Q4_K_M
Context length is an Ollama server setting: quit the Ollama app or stop the service first, or set it in its environment (launchctl on Mac, systemctl on Linux).
Best for Vision · 16k context
| Model | Quant | Fit | Memory | Speed (generation) |
|---|---|---|---|---|
| Nemotron 3.5 Super VL 120B-A12B | Q4_K_M | fits fully | 70.4 GB | ~25 tok/s estimated |
| Qwen3.8 Flash Next | UD-Q5_K_XL | fits fully | 158.8 GB | ~33 tok/s estimated |
| Gemma 4 12B | Q8_0 | fits fully | 13.3 GB | ~17 tok/s estimated |
| Qwen3.6 35B-A3B | Q8_0 | fits fully | 37.3 GB | ~56 tok/s estimated |
| Gemma 4 26B-A4B | Q8_0 | fits fully | 27.4 GB | ~43 tok/s estimated |
llama-server -hf lmstudio-community/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-GGUF:Q4_K_M -c 16384 -fa on --jinja -ngl 99
OLLAMA_CONTEXT_LENGTH=16384 ollama serve ollama run hf.co/lmstudio-community/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-GGUF:Q4_K_M
Context length is an Ollama server setting: quit the Ollama app or stop the service first, or set it in its environment (launchctl on Mac, systemctl on Linux).
All 28 models that fit, fastest first · 16k context
Click a column header to sort. 0 of 28 speeds here are measured or scaled from measured; the rest are bandwidth estimates. Filter by use case, speed source, fit, license and more in the interactive advisor →
| Model | Params | Quant | Fit | Memory | Tok/s | Source |
|---|---|---|---|---|---|---|
| gpt-oss 20B | 20.91B MoE | Q8_0 | fits fully | 12.5 GB | ~82 | estimated |
| LFM2.5 24B-A2B | 23.97B MoE | Q8_0 | fits fully | 25.8 GB | ~66 | estimated |
| gpt-oss 120B | 116.83B MoE | Q8_0 | fits fully | 64.0 GB | ~62 | estimated |
| GLM-4.7-Flash | 31.22B MoE | Q8_0 | fits fully | 32.7 GB | ~56 | estimated |
| Qwen3.6 35B-A3B | 35.95B MoE | Q8_0 | fits fully | 37.3 GB | ~56 | estimated |
| Ornith 1.5 35B-A3B | 35.95B MoE | Q8_0 | fits fully | 38.2 GB | ~54 | estimated |
| Laguna XS 2.1 | 33.44B MoE | Q8_0 | fits fully | 36.3 GB | ~54 | estimated |
| Nemotron 3.5 Lightning 30B-A3B | 31.58B MoE | Q8_0 | fits fully | 35.2 GB | ~52 | estimated |
| Nemotron 3 Nano 4B | 3.97B | Q8_0 | fits fully | 4.6 GB | ~52 | estimated |
| Qwen3-Coder 30B-A3B | 30.53B MoE | Q8_0 | fits fully | 34.1 GB | ~49 | estimated |
| Qwen3.5 4B | 4.66B | Q8_0 | fits fully | 5.1 GB | ~49 | estimated |
| Nemotron 3 Nano 30B-A3B | 31.58B MoE | Q8_0 | fits fully | 33.7 GB | ~46 | estimated |
| Gemma 4 26B-A4B | 25.81B MoE | Q8_0 | fits fully | 27.4 GB | ~43 | estimated |
| Qwen3.8 Flash Next | 180B MoE | UD-Q5_K_XL | fits fully | 158.8 GB | ~33 | estimated |
| DeepSeek R1 Distill Qwen 7B | 7.62B | Q8_0 | fits fully | 9.0 GB | ~27 | estimated |
| Gemma 4 E4B | 8B | Q8_0 | fits fully | 8.5 GB | ~27 | estimated |
| Nemotron 3.5 Super VL 120B-A12B | 124.45B MoE | Q4_K_M | fits fully | 70.4 GB | ~25 | estimated |
| Qwen3-VL 8B | 8.77B | Q8_0 | fits fully | 11.1 GB | ~25 | estimated |
| Qwen3.5 9B | 9.65B | Q8_0 | fits fully | 10.1 GB | ~23 | estimated |
| Ornith 1.5 9B | 9.65B | Q8_0 | fits fully | 10.1 GB | ~23 | estimated |
| Gemma 4 12B | 11.96B | Q8_0 | fits fully | 13.3 GB | ~17 | estimated |
| GLM-4.5-Air | 110.47B MoE | Q6_K | fits fully | 102.1 GB | ~16 | estimated |
| Mistral Small 3.2 24B | 24.01B | Q4_K_M | fits fully | 17.0 GB | ~15 | estimated |
| Qwen3.8 27B | 27.78B | UD-Q4_K_M | fits fully | 17.7 GB | ~13 | estimated |
| Qwen3.6 27B | 27.78B | Q4_K_M | fits fully | 18.1 GB | ~13 | estimated |
| Gemma 4 31B | 31.27B | Q4_K_M | fits fully | 20.5 GB | ~12 | estimated |
| DeepSeek R1 Distill Qwen 32B | 32.76B | Q4_K_M | fits fully | 24.1 GB | ~11 | estimated |
| Llama 3.3 70B | 70.55B | Q4_K_M | fits fully | 47.9 GB | ~5 | estimated |
Long context on this hardware · up to 1M tokens
Largest context whose KV cache still fits next to the weights (same placement as the 16k result, KV types f16 / q8_0 / q4_0). Models are capped at their verified maximum; extended means beyond the native maximum using YaRN, only listed where the model card documents it. Long context needs flash attention (-fa on, already in the commands) and a lot of memory; a quantized KV cache saves memory but can cost accuracy. Speed at depth is an estimate (penalty fitted to measured depth runs up to 128k).
Try it in the advisor: 128k · 256k · 512k · 1M
| Model | Native max | Max ctx f16 | Max ctx q8_0 | Max ctx q4_0 | Est. tok/s at max f16 ctx |
|---|---|---|---|---|---|
| Qwen3.6 35B-A3B Q8_0 | 256k | 1M extended | 1M extended | 1M extended | ~9 estimated |
| Qwen3-Coder 30B-A3B Q8_0 | 256k | 1M extended | 1M extended | 1M extended | ~2 estimated |
| Qwen3.5 4B Q8_0 | 256k | 1M extended | 1M extended | 1M extended | ~8 estimated |
| Nemotron 3.5 Super VL 120B-A12B Q4_K_M | 1M | 1M | 1M | 1M | ~13 estimated |
| Qwen3-VL 8B Q8_0 | 256k | 1M extended | 1M extended | 1M extended | ~2 estimated |
| Qwen3.5 9B Q8_0 | 256k | 1M extended | 1M extended | 1M extended | ~6 estimated |
| Qwen3.8 27B UD-Q4_K_M | 256k | 1M extended | 1M extended | 1M extended | ~3 estimated |
| Qwen3.6 27B Q4_K_M | 256k | 1M extended | 1M extended | 1M extended | ~3 estimated |
| Ornith 1.5 35B-A3B Q8_0 | 256k | 256k | 256k | 256k | ~24 estimated |
| Laguna XS 2.1 Q8_0 | 256k | 256k | 256k | 256k | ~15 estimated |
| Nemotron 3.5 Lightning 30B-A3B Q8_0 | 256k | 256k | 256k | 256k | ~38 estimated |
| Nemotron 3 Nano 4B Q8_0 | 256k | 256k | 256k | 256k | ~29 estimated |
| Nemotron 3 Nano 30B-A3B Q8_0 | 256k | 256k | 256k | 256k | ~35 estimated |
| Gemma 4 26B-A4B Q8_0 | 256k | 256k | 256k | 256k | ~21 estimated |
| Ornith 1.5 9B Q8_0 | 256k | 256k | 256k | 256k | ~14 estimated |
| Gemma 4 12B Q8_0 | 256k | 256k | 256k | 256k | ~14 estimated |
| Gemma 4 31B Q4_K_M | 256k | 256k | 256k | 256k | ~6 estimated |
| gpt-oss 20B Q8_0 | 128k | 128k | 128k | 128k | ~38 estimated |
| LFM2.5 24B-A2B Q8_0 | 128k | 128k | 128k | 128k | ~37 estimated |
| gpt-oss 120B Q8_0 | 128k | 128k | 128k | 128k | ~27 estimated |
| GLM-4.7-Flash Q8_0 | 198k | 128k | 128k | 128k | ~21 estimated |
| Qwen3.8 Flash Next UD-Q5_K_XL | 256k | 128k | 256k | 256k | ~22 estimated |
| DeepSeek R1 Distill Qwen 7B Q8_0 | 128k | 128k | 128k | 128k | ~16 estimated |
| Gemma 4 E4B Q8_0 | 128k | 128k | 128k | 128k | ~22 estimated |
| GLM-4.5-Air Q6_K | 128k | 128k | 128k | 128k | ~6 estimated |
| Mistral Small 3.2 24B Q4_K_M | 128k | 128k | 128k | 128k | ~7 estimated |
| DeepSeek R1 Distill Qwen 32B Q4_K_M | 128k | 128k | 128k | 128k | ~5 estimated |
| Llama 3.3 70B Q4_K_M | 128k | 128k | 128k | 128k | ~3 estimated |
FAQ
What is the best local LLM for AMD Ryzen AI Max+ PRO 495 (Ryzen AI Halo) 192GB?
Of the 28 models that fit fully at 16k context, gpt-oss 20B Q8_0 is the fastest at ~82 tok/s (estimated); the largest is Qwen3.8 Flash Next UD-Q5_K_XL.
How much memory can AMD Ryzen AI Max+ PRO 495 (Ryzen AI Halo) 192GB use for a model?
About 162.7 GB usable for the model and KV cache, with 273.1 GB/s memory bandwidth.
Can AMD Ryzen AI Max+ PRO 495 (Ryzen AI Halo) 192GB run long contexts?
See the long-context table above: it lists the largest context whose KV cache fits next to the weights for every model, up to 1M tokens where the model supports it.
Open the interactive advisor with this hardware →
Speeds marked measured are medians of real runs (short-context generation) from these sources; scaled from measured means a measurement of a same-size or same-model-different-quant run, scaled by bytes read; estimated comes from memory bandwidth (spec source: https://www.amd.com/en/products/processors/laptop/ryzen-pro/ai-max-pro-400-series/amd-ryzen-ai-max-plus-pro-495.html) × an efficiency factor of 0.80 calibrated to measured runs on this hardware class. No measured data on file for this hardware yet · submit your benchmark. Memory size on this page is an example configuration; check what your machine has. Model data last verified 2026-10-20. Ranking = fits and fast enough first, then larger; not a quality benchmark.
Page generated 2026-10-21 by tools/generate_pages.js