Qwen3.6-35B-A3B
Qwen3.6-35B-A3B is a Qwen MoE text model (~35B total / 3B active). This chapter has two hardware paths: Lemonade / llama.cpp on GGUF + iGPU, FastFlowLM on NPU2 + XDNA2. Do not feed GGUF to flm, or model.q4nx to llama.cpp.
GPU tutorials for earlier Qwen generations: Qwen3 · Qwen3.5. A Lemonade id like
Qwen3.5-35B-A3B-GGUFis a different generation — use an id that contains 3.6.
| Path | Format | Typical size | Runtime |
|---|---|---|---|
| GPU / Lemonade | GGUF (Qwen3.6-35B-A3B-GGUF, often UD-Q4) | ~20–23 GiB | Lemonade llamacpp:vulkan (or rocm) @ 13305 |
| NPU / FastFlowLM | NPU2 | ~29 GiB resident | flm serve qwen3.6-moe:35b-a3b @ 8219 |
This is hungrier than Gemma 4 E4B (~9 GiB). Stop other large models first. On the validation host, NPU short-prompt concurrency 1/2/4/8 passed; long-form decode was about 14 tok/s.
Deploy: FastFlowLM NPU · Lemonade GPU · Lemonade NPU