Can I run it? / Ryzen AI Max+ 395 (Strix Halo)
Best LLMs to run on a Ryzen AI Max+ 395 (Strix Halo)
Updated November 2026
As of November 2026, the best model to run on a Ryzen AI Max+ 395 (Strix Halo) is Qwen3.8 27B by Alibaba (Qwen) (Small LM Index 74.6), at about 9 tokens/s with 4-bit weights. For faster replies, Kolibri-1 (71.1) runs at about 37 tokens/s.
Strix Halo mini PCs and laptops share up to 128 GB between CPU and GPU at 256 GB/s: room for 70B-class models, at speeds best suited to Mixture-of-Experts models. About 96 GB is usable for a model, at 256 GB/s. The figures below assume 4-bit weights and a 8K-token context, and only include models that generate at least 5 tokens per second.
Models that run well on a Ryzen AI Max+ 395 (Strix Halo)
The 25 best of the 101 models that run at 5+ tokens/s, ranked by the Small LM Index.
| # | Model | Index | Memory | Speed |
|---|---|---|---|---|
| 1 | Qwen3.8 27B Alibaba (Qwen) · 4-bit · Apache-2.0 | 74.6 | 18 GB | ≈ 9 tok/s |
| 2 | Nemotron 3.5 Bolt 74B-A9B NVIDIA · 4-bit · OpenMDW-1.1 | 73.4 | 47 GB | ≈ 14 tok/s |
| 3 | Ternary Bonsai 2 27B PrismML · Q2_0 (as released) · Apache-2.0 | 71.2 | 9.3 GB | ≈ 20 tok/s |
| 4 | Kolibri-1 Aleph Alpha · 4-bit · Apache-2.0 | 71.1 | 50 GB | ≈ 37 tok/s |
| 5 | Nanbeige4.3 33B Nanbeige (Kanzhun) · 4-bit · Apache-2.0 | 70.9 | 23 GB | ≈ 8 tok/s |
| 6 | Muse Glimmer 30B Meta · 4-bit · Apache-2.0 | 70.1 | 19 GB | ≈ 9 tok/s |
| 7 | Qwen3.6 27B Alibaba (Qwen) · 4-bit · Apache-2.0 | 69.7 | 18 GB | ≈ 9 tok/s |
| 8 | Qwen3.6 35B-A3B Alibaba (Qwen) · 4-bit · Apache-2.0 | 68.3 | 23 GB | ≈ 42 tok/s |
| 9 | Qwen3.5 27B Alibaba (Qwen) · 4-bit · Apache-2.0 | 68.3 | 18 GB | ≈ 9 tok/s |
| 10 | Nanbeige4.2 3B Nanbeige (Kanzhun) · 4-bit · Apache-2.0 | 67.2 | 3.8 GB | ≈ 61 tok/s |
| 11 | Qwen3.5 35B-A3B Alibaba (Qwen) · 4-bit · Apache-2.0 | 66.9 | 23 GB | ≈ 42 tok/s |
| 12 | Spark-X2.6 4B SparkLLM (XHToken) · 4-bit · Apache-2.0 | 66.5 | 3.4 GB | ≈ 62 tok/s |
| 13 | Gemma 4 31B Google · 4-bit · Apache-2.0 | 66.2 | 22 GB | ≈ 8 tok/s |
| 14 | GLM-4.7-Flash Z.ai (GLM) · 4-bit · MIT | 66.0 | 20 GB | ≈ 42 tok/s |
| 15 | Spark-X2.5 4B SparkLLM (XHToken) · 4-bit · Apache-2.0 | 65.9 | 3.4 GB | ≈ 62 tok/s |
| 16 | Granite 4.2 30B IBM (Granite) · 4-bit · Apache-2.0 | 65.8 | 21 GB | ≈ 8 tok/s |
| 17 | Qwen3 30B-A3B Thinking 2507 Alibaba (Qwen) · 4-bit · Apache-2.0 | 65.4 | 20 GB | ≈ 38 tok/s |
| 18 | Gemma 4 26B-A4B Google · 4-bit · Apache-2.0 | 64.8 | 17 GB | ≈ 33 tok/s |
| 19 | North Mini Code Cohere · 4-bit · Unverified | 64.7 | 21 GB | ≈ 42 tok/s |
| 20 | gpt-oss-20b OpenAI · 4-bit · Apache-2.0 | 64.7 | 14 GB | ≈ 35 tok/s |
| 21 | Nemotron 3 Nano 30B-A3B NVIDIA · 4-bit · NVIDIA Open Model | 64.4 | 20 GB | ≈ 36 tok/s |
| 22 | Qwen3.5 9B Alibaba (Qwen) · 4-bit · Apache-2.0 | 64.1 | 6.4 GB | ≈ 28 tok/s |
| 23 | Qwen3-Next 80B-A3B Thinking Alibaba (Qwen) · 4-bit · Apache-2.0 | 63.7 | 51 GB | ≈ 42 tok/s |
| 24 | Nemotron 3.5 Lightning NVIDIA · 4-bit · OpenMDW-1.1 | 63.3 | 22 GB | ≈ 35 tok/s |
| 25 | Apriel 1.6 15B Thinker ServiceNow (Apriel) · 4-bit · MIT | 63.3 | 11 GB | ≈ 17 tok/s |
Change the context length or quantization, or see models that only just fit, in Can I run it? for this device.
How to run Qwen3.8 27B on a Ryzen AI Max+ 395 (Strix Halo)
- Ollama, LM Studio or llama.cpp. All three use the integrated GPU (Vulkan or ROCm on Ryzen, CUDA on DGX Spark).
ollama run hf.co/ggml-org/Qwen3.8-27B-GGUF:Q4_K_M
- Give the GPU more memory. On Ryzen AI Max systems, raise the GPU memory allocation in the BIOS or driver so large models fit.
Every model page lists the same commands for its own files. Qwen3.8 27B's page also has a memory table for other context lengths.
Questions
What is the best LLM for a Ryzen AI Max+ 395 (Strix Halo)?
As of November 2026, the best model to run on a Ryzen AI Max+ 395 (Strix Halo) is Qwen3.8 27B by Alibaba (Qwen) (Small LM Index 74.6), at about 9 tokens/s with 4-bit weights. For faster replies, Kolibri-1 (71.1) runs at about 37 tokens/s.
What is the biggest model that runs on a Ryzen AI Max+ 395 (Strix Halo)?
Qwen3-Next 80B-A3B Thinking (80B parameters, 3B active) fits at 4-bit, needing 51 GB and generating about 42 tokens/s. Bigger models may fit at lower quality or run more slowly; try them in the Can I run it? tool.
How much memory can a model use on a Ryzen AI Max+ 395 (Strix Halo)?
About 96 GB of the 128 GB. Strix Halo mini PCs and laptops share up to 128 GB between CPU and GPU at 256 GB/s: room for 70B-class models, at speeds best suited to Mixture-of-Experts models.
How accurate are these speeds?
They are estimates from memory bandwidth (256 GB/s here), which limits how fast a model generates text, with a typical efficiency for this kind of hardware. Expect real speeds within about a third either way. Long prompts take longer to read before the first word appears.