Best local LLM for AMD Strix Halo mini PCs and AI boxes.
These small boxes share up to 128 GB or more between the CPU and graphics, like a Mac. They fit big models for less money, but are usually slower per token than a Mac Studio.
Published speeds.
Other people's measurements, with sources. Not our tests.
| Machine | Model | Tokens/sec | Engine | Source |
|---|---|---|---|---|
| NVIDIA DGX Spark (128GB) | gpt-oss-120b | 55.37 | llama.cpp · MXFP4 | 2025-10-24 |
| NVIDIA DGX Spark (128GB) | gpt-oss-20b | 82.74 | llama.cpp · MXFP4 | 2025-10-24 |
| Framework Desktop (Ryzen AI Max+ 395) | gpt-oss-120b | 54 | llama.cpp Vulkan · MXFP4 (GGUF) | 2026-10-15 |
| GMKtec EVO-X2 (Ryzen AI Max+ 395) | gpt-oss-120b | 46.05 | llama.cpp ROCm (llama-bench) · MXFP4 (GGUF) | 2026-10-15 |
| Minisforum MS-S1 MAX (Ryzen AI Max+ 395) | gpt-oss-20b | 75.7 | llama.cpp ROCm 7.1 · UD-Q4_K_XL | 2026-10-15 |
| Minisforum MS-S1 MAX-P495 (Ryzen AI Max+ PRO 495 "Gorgon Halo") | qwen3.5-397b-a17b | 16 | ? · | 2026-10-16 |
Short answers.
- What is the best local LLM for the smallest Mini PC?
- With 64 GB (DGX Spark 64GB (partner GB10 systems)), the best fit is Qwen3.8 27B (needs 22.7 GB). 77 models fit.
- Which Mini PC runs the smartest local model?
- Gigabyte W775-V10-L01 (DGX Station GB300) with 252 GB runs GLM-5.3-Flash (AA 41.8, about 73% of Claude Opus 5.5).
- What matters more, memory or speed?
- Memory decides which models fit at all. Bandwidth (GB/s) decides how fast they answer. Pick memory first, then the fastest machine you can afford with that memory.
Scores are from Artificial Analysis and LMArena, not our own tests yet. Memory need = 4-bit file × 1.15 + 2 GB; a model fits when that is at most 75% of memory. Models 2026-10-15, machines 2026-10-15.
# Best local LLM for AMD Strix Halo mini PCs and AI boxes URL: https://insidethemachine.ai/best/local-llm-for-mini-pc Data: machines 2026-10-15, models 2026-10-15. These small boxes share up to 128 GB or more between the CPU and graphics, like a Mac. They fit big models for less money, but are usually slower per token than a Mac Studio. | machine | memory GB | usable GB | bandwidth GB/s | models that fit | best fit | needs GB | |---|---|---|---|---|---|---| | NVIDIA DGX Spark (128GB) | 128 | 96 | 273 | 90 | Qwen3.8 27B | 22.7 | | DGX Spark 64GB (partner GB10 systems) | 64 | 48 | 273 | 77 | Qwen3.8 27B | 22.7 | | Framework Desktop (Ryzen AI Max+ 395) | 64 | 48 | 256 | 77 | Qwen3.8 27B | 22.7 | | Framework Desktop (Ryzen AI Max+ 395) | 128 | 96 | 256 | 90 | Qwen3.8 27B | 22.7 | | GMKtec EVO-X2 (Ryzen AI Max+ 395) | 64 | 48 | 256 | 77 | Qwen3.8 27B | 22.7 | | GMKtec EVO-X2 (Ryzen AI Max+ 395) | 128 | 96 | 256 | 90 | Qwen3.8 27B | 22.7 | | Minisforum MS-S1 MAX (Ryzen AI Max+ 395) | 64 | 48 | 256 | 77 | Qwen3.8 27B | 22.7 | | Minisforum MS-S1 MAX (Ryzen AI Max+ 395) | 128 | 96 | 256 | 90 | Qwen3.8 27B | 22.7 | | Bosgame M5 (Ryzen AI Max+ 395) | 128 | 96 | 256 | 90 | Qwen3.8 27B | 22.7 | | Beelink GTR9 Pro (Ryzen AI Max+ 395) | 128 | 96 | 256 | 90 | Qwen3.8 27B | 22.7 | | GMKtec EVO-X5 Pro (Ryzen AI Max+ PRO 495 "Gorgon Halo") | 192 | 159.9 | 273 | 94 | Qwen3.8-Flash-Next (125B-A6B) | 140 | | Minisforum MS-S1 MAX-P495 (Ryzen AI Max+ PRO 495 "Gorgon Halo") | 192 | 159.9 | ? | 94 | Qwen3.8-Flash-Next (125B-A6B) | 140 | | Gigabyte W775-V10-L01 (DGX Station GB300) | 252 | 251 | 7100 | 97 | GLM-5.3-Flash | 233 | | ASUS ProArt GR1X (NVIDIA RTX Spark) | 64 | 48 | ? | 77 | Qwen3.8 27B | 22.7 | | ASUS ProArt GR1X (NVIDIA RTX Spark) | 128 | 96 | ? | 90 | Qwen3.8 27B | 22.7 | | iBUYPOWER Pro AMD AI Halo Mini PC (Ryzen AI Max+ PRO 495) | 128 | 96 | ? | 90 | Qwen3.8 27B | 22.7 | | Microsoft Surface RTX Spark Dev Box (128GB) | 128 | 96 | ? | 90 | Qwen3.8 27B | 22.7 | | AMD Ryzen AI Halo Developer Platform (Ryzen AI Max+ PRO 495, 192GB) | 192 | 159.9 | ? | 94 | Qwen3.8-Flash-Next (125B-A6B) | 140 | ## Published speeds (decode tokens/sec, not our tests) - NVIDIA DGX Spark (128GB): gpt-oss-120b 55.37 tok/s (llama.cpp, MXFP4, ) source: https://developer.nvidia.com/blog/how-nvidia-dgx-sparks-performance-enables-intensive-ai-tasks/ (2025-10-24) - NVIDIA DGX Spark (128GB): gpt-oss-20b 82.74 tok/s (llama.cpp, MXFP4, ) source: https://developer.nvidia.com/blog/how-nvidia-dgx-sparks-performance-enables-intensive-ai-tasks/ (2025-10-24) - Framework Desktop (Ryzen AI Max+ 395): gpt-oss-120b 54 tok/s (llama.cpp Vulkan, MXFP4 (GGUF), Framework Desktop 128GB (llama-bench pp512/tg128)) source: https://llamaperf.com/model/gpt-oss (2026-10-15) - GMKtec EVO-X2 (Ryzen AI Max+ 395): gpt-oss-120b 46.05 tok/s (llama.cpp ROCm (llama-bench), MXFP4 (GGUF), EVO-X2 128GB, Ubuntu 24.04 (Nish Tahir, 2025-10-23)) source: https://llamaperf.com/model/gpt-oss (2026-10-15) - Minisforum MS-S1 MAX (Ryzen AI Max+ 395): gpt-oss-20b 75.7 tok/s (llama.cpp ROCm 7.1, UD-Q4_K_XL, ) source: https://llamaperf.com/model/gpt-oss (2026-10-15) - Minisforum MS-S1 MAX-P495 (Ryzen AI Max+ PRO 495 "Gorgon Halo"): qwen3.5-397b-a17b 16 tok/s (engine?, quant?, two 192GB units clustered) source: https://store.minisforum.com/products/minisforum-ms-s1-max-p495-ai-workstation (2026-10-16) ## Questions **What is the best local LLM for the smallest Mini PC?** With 64 GB (DGX Spark 64GB (partner GB10 systems)), the best fit is Qwen3.8 27B (needs 22.7 GB). 77 models fit. **Which Mini PC runs the smartest local model?** Gigabyte W775-V10-L01 (DGX Station GB300) with 252 GB runs GLM-5.3-Flash (AA 41.8, about 73% of Claude Opus 5.5). **What matters more, memory or speed?** Memory decides which models fit at all. Bandwidth (GB/s) decides how fast they answer. Pick memory first, then the fastest machine you can afford with that memory.