Chat on the AMD Ryzen AI NPU works once a user finds it. On a Strix Halo machine, though, the NPU section in the model picker would not open from its heading, the CLI could not use NPU models, and one common failure showed a blank reply. This fixes those gaps.

Changes

Chat

• When a reply's reasoning fills the NPU context, the chat now says "Response filled the model's context window. Start a new chat, or shorten this one, to keep going." Before, it showed only "Worked for 2m 37s", because FastFlowLM ends that stream with stop instead of length. API clients now get length too.
• Collapsible sections in the model picker now open from their label, not only from the 12 px chevron at the far edge. This applies to every collapsible section, including the NPU one that holds Enable NPU. The NPU setup text is also plainer.

CLI

• unsloth run --model qwen3-0.6b-FLM (or lemonade:<id>) enables the NPU runtime, downloads the model, and serves it. Before, the id was parsed as a GGUF repo. For NPU models, the banner shows only the chat example.
• unsloth chat qwen3-0.6b-FLM chats through a running Studio. Without one, it says how to start it instead of crashing.
• unsloth start explains that coding agents need a GGUF model instead of looking the id up on Hugging Face.

API

• With "Switch model by request" on, naming a downloaded NPU model loads it under the same safeguards as a GGUF switch; a cancelled request does not switch. Requests the NPU model could not serve are refused before anything is unloaded: Anthropic Messages, streaming Responses, audio or video input, and images, including tool-result images, for a model without vision.
• Loading an NPU model while another model is serving now waits for active inference first, as GGUF and transformers loads do.
• /v1/messages with an NPU model loaded explains that NPU models serve /v1/chat/completions only. Before, it said "No GGUF model loaded".
• Tool-call replies no longer carry "\n\n\n" as content.

Limits

• Coding agents, Anthropic Messages, completions and embeddings still need a GGUF model.
• A reasoning-only reply that does not fill the context still saves without a notice. That rule is shared by every provider.

Screenshots

Model picker after clicking the "NPU" heading (two isolated installs, NPU responses mocked):

A reply whose reasoning fills a 2048-token context, on the Strix Halo machine:

Verification

On AMD Ryzen AI MAX+ 395 machines, main (A) and this branch (B) were installed from source on Linux and on Windows, and both ran the same steps:

On B, with switching on, requests the NPU model cannot serve were refused without unloading anything (Linux and Windows). On Linux, with a GGUF reply streaming on the iGPU, a request naming the NPU model waited for the reply to finish before loading.

New backend and CLI tests fail without these fixes. The related backend, CLI and frontend suites show no new failures, and the type check is clean.

Before / after (merge base 5e927b0 vs head f39640f)

Two isolated Studio installs, one per commit. The NPU API is mocked in the browser, as in the screenshots above.

Clicking the "NPU" heading text in the model picker. Before, the section stays folded. After, it opens:

The NPU section opened. The setup text is plainer, and the "Powered by FastFlowLM ... and Lemonade" line is gone. It is also gone from the Hub's NPU list:

Every collapsible section opens from its label, not only from the chevron. Here, "Other models" on the On Device tab:

Follow-up changes

• The "Powered by FastFlowLM and Lemonade" credit line is removed from the NPU setup notice and the Hub NPU list.
• A full NPU context is now read from active_kv_tokens. total_tokens came up 3 short when thinking was off, because the empty think block counts four context tokens as one generated token, so the context-window notice did not show.
• A request with an MCP image that names an NPU model is now refused before the switch. Before, it unloaded the serving model first.
• The NPU is off by default. Start Studio with UNSLOTH_STUDIO_ENABLE_NPU=1 to use it. Without the flag, an XDNA2 NPU reports supported: false, so the picker and Hub show no NPU section, and NPU loads, the runtime start, and unsloth run / unsloth chat with an NPU id are refused with a message naming the flag. The screenshots above were taken with the NPU on.