Boards / M5Stack LLM8850

M5Stack LLM8850 M.2 (AX8850): LLM benchmarks

An AX8850 accelerator on an M.2 card, for adding an NPU to another host.

At a glance

  • Measured runs7
  • Models tested7
  • Fastest run27.05 tok/s
  • Qwen2.5 0.5B Instruct, W4A16, StackFlow · source ↗
  • Sources1

Where to buy

Every run on this board

ModelQuantRuntimeGen tok/sPrompt tok/sTTFT msSource
Qwen2.5 0.5B InstructW4A16StackFlow27.05––source ↗
accessed 2026-10-17
MiniCPM4 0.5BW8A16StackFlow21.05–212.91source ↗
accessed 2026-10-17
Qwen2.5 1.5B InstructW4A16StackFlow15.06––source ↗
accessed 2026-10-17
DeepSeek-R1-Distill-Qwen 1.5BW4A16StackFlow13.29––source ↗
accessed 2026-10-17
Qwen3 0.6BW8A16StackFlow12.88–670.51source ↗
accessed 2026-10-17
Qwen3-VL 2B InstructW8A16StackFlow7.8––source ↗
accessed 2026-10-17
Qwen3 1.7BW8A16StackFlow7.38–796.38source ↗
accessed 2026-10-17

Can the M5Stack LLM8850 run…

Can it run DeepSeek-R1-Distill-Qwen 1.5B?Can it run MiniCPM4 0.5B?Can it run Qwen2.5 0.5B?Can it run Qwen2.5 1.5B?Can it run Qwen3 0.6B?Can it run Qwen3 1.7B?Can it run Qwen3-VL 2B?

Have this board? Submit your own run →

Get the monthly board report

New boards, new runtimes, new numbers. One email a month, no hype.