Since 2026.10.7 the NPU section is gone from the model picker and the Hub on every Ryzen AI machine, including installs that were already chatting on the NPU. #13797 put it behind UNSLOTH_STUDIO_ENABLE_NPU=1, so /api/npu/status reports supported: false on an XDNA2 NPU unless that variable is set, and nothing in the app says so.

This brings the NPU back by default and makes the first-run notice clearer.

Changes

Backend and CLI

• An XDNA2 NPU is shown again with the variable unset. UNSLOTH_STUDIO_ENABLE_NPU=0 (or false, no, off) still hides it, and NPU loads, the runtime start and unsloth run / unsloth chat with an NPU id are refused with a message naming the variable.
• XDNA1 stays unsupported, as before.

Setup notice (model picker NPU section and the Hub NPU list)

• Before the NPU is enabled, the notice is a short card: "This PC has an AMD Ryzen AI NPU", one sentence saying the NPU runs its own builds of models such as Qwen and Gemma alongside the GPU models, and "One-time setup: downloads about 47 MB and checks the NPU driver." next to the Enable NPU button. The wording reads the same in the model picker and in the Hub.
• Once enabled, the picker shows one line: "These models run on the AMD NPU, not on the GPU."
• The runtime names are gone from the notice. The "Powered by FastFlowLM and Lemonade" credit line was already removed in Studio: improve AMD NPU discovery, CLI support, and chat feedback #13797 and stays removed.
• Errors, the driver setup link, Try again and Start NPU runtime behave as before.

Testing

• studio/backend/tests/test_npu_detect.py, test_npu_backend.py, test_npu_chat_route.py, test_flm_files.py and unsloth_cli/tests/test_npu_cli.py pass. The detection test now covers unset, empty, 1, 0, false and Off; on main the unset and empty cases fail, since main hides the NPU there.
• test_export_npu_q4nx.py::test_missing_converter_deps_fail_before_installing fails on main too and is unrelated.
• Frontend type check is clean and the NPU frontend tests pass.

Screenshots

Two isolated installs, main (ab26545) and this branch. This build host has no NPU, so both Studios were started with the backend's Linux PCI probe reporting a Strix Halo XDNA2 NPU; the gate, /api/npu/status and the frontend are each side's own code. Nothing was enabled or downloaded.

Model picker, Recommended tab, NPU section opened, with UNSLOTH_STUDIO_ENABLE_NPU unset. Main shows no NPU section at all; this branch shows it with the new card:

The same with UNSLOTH_STUDIO_ENABLE_NPU=1 on both sides, so main's notice is visible too. The runtime names are gone, and the card says what the NPU is for and what Enable does: