Castellan
Supported NPUs
The NPUs Castellan runs on.
Castellan's local AI runs on the NPU in these chips. We say exactly which chips we've tested, which we expect to work from those tests, and which are untested, with our benchmark of each chip we tested beside it.
- Tested
- We ran Castellan's NPU check on this exact chip, in a real laptop: it installs the model server and the model, then tests and times every step, and it passed them all. Then we benchmarked it.
- Expected
- We haven't run these chips. Each has the same NPU as a chip we tested, or a newer one whose settings we've estimated from them.
- Untested
- The setup supports these, but we haven't run them, or anything with their NPU, so we can't say how well they do.
Whichever it is, the setup asks your NPU a test question and times it before it's used. If it doesn't answer, your graphics card or processor does the work.
| Chip | Status | Job time | Writing | Short answer | Start | Answers right |
|---|---|---|---|---|---|---|
| Tested: the exact chips we ran | ||||||
Snapdragon X2 Elite Extreme (X2E94100) Hexagon NPU, 80 TOPS Qwen3-VL-4B Instruct, 4-bit, on GenieX | Tested In our ASUS Zenbook A16, 7 October 2026 | — | 34 tokens/s | — | 9–15 s | — |
Intel Core Ultra 7 256V NPU 4 Qwen3-4B Instruct 2507, 4-bit, on OpenVINO Model Server 2026.4 | Tested In our HP OmniBook X 17, 12 October 2026 | 11.85 s | 25.7 tokens/s | 0.82 s | 4.5 s | 23 of 24 |
Intel Core Ultra 5 225U NPU 3720 Qwen3-4B Instruct 2507, 4-bit, on OpenVINO Model Server 2026.4 | Tested In our Lenovo IdeaPad 5i, 14 October 2026 | 34.5 s | 9.4 tokens/s | 3.1 s | 6.9 s | 23 of 24 |
AMD Ryzen AI 7 350 XDNA 2 NPU, 50 TOPS Qwen3-4B Instruct 2507, 4-bit, on FastFlowLM 1.0.7 | Tested In our test PC with 16 GB of memory, 15 October 2026 | — | 13.8 tokens/s | — | 6.6 s | — |
| Expected to work: estimated from our tests | ||||||
Snapdragon X2 Elite and X2 Plus, the other models The same Hexagon NPU generation | Expected The same NPU as the X2 Elite Extreme we tested, on its settings. | Not measured | ||||
Intel Core Ultra 200V, the other models NPU 4, up to 48 TOPS | Expected The same NPU as the Core Ultra 7 256V we tested, on its settings. | Not measured | ||||
Intel Core Ultra Series 1, and the other 200U models NPU 3720 | Expected The same NPU as the Core Ultra 5 225U we tested, on its settings. | Not measured | ||||
Intel Core Ultra 200S (desktop) NPU 3720 | Expected The same NPU as the Core Ultra 5 225U we tested, on its settings. | Not measured | ||||
Intel Core Ultra Series 3 NPU 5, up to 50 TOPS | Expected A newer NPU we haven't run, about as fast as the Core Ultra 7 256V's: its settings are estimated from that one's. | Not measured | ||||
Intel Core Series 3 NPU 5, up to 17 TOPS | Expected A newer NPU we haven't run, nearer the Core Ultra 5 225U's in speed: its settings are estimated from the Intel chips we tested. | Not measured | ||||
AMD Ryzen AI 300, Ryzen AI Max and Ryzen AI 400, the other models XDNA 2 NPU | Expected The same NPU as the Ryzen AI 7 350 we tested, on its settings. | Not measured | ||||
| Untested | ||||||
Snapdragon X Elite and X Plus (the first generation) Hexagon NPU, 45 TOPS | Untested We haven't run it, or any chip with its NPU: the setup tries it on the X2 Elite's settings. | Not measured | ||||
- Snapdragon X2 Elite Extreme (X2E94100): From our tests before the benchmark: the full benchmark hasn't run here yet.
- Intel Core Ultra 5 225U: The older NPU 3720, as in Core Ultra Series 1 and 200U chips. Measured with OpenVINO's own prompt settings, as the 256V was. With the settings Castellan now sets up for this NPU, tuned on this laptop, the benchmark's quick run did the jobs in 23 s, wrote 11 tokens a second and gave a short answer in 0.5 s (14 October 2026).
- AMD Ryzen AI 7 350: From its NPU check, on mains power: the full benchmark hasn't run here yet.
Only the chips we tested have figures: we don't estimate a speed we haven't measured. The benchmark is ours, on our own laptops, plugged in, in Windows' default power mode. Job time is three typical pieces of the staff's work back to back: sorting an email thread, drafting a reply and summarising a document. Lower is better. A token is about three-quarters of a word. Short answer is the wait for a one-word reply. Start is how long a stopped model takes to answer again. Answers right is out of the same 24 graded questions. A dash is a figure we haven't measured yet. The graphics card beside one of these NPUs, for comparison, is in How fast it runs.
Not on the list?
Your PC still gets its local AI.
On a newer Snapdragon X, Intel Core Ultra or AMD Ryzen AI NPU the setup doesn't know yet, it tries the NPU on cautious settings and keeps it only if it answers the test question. It can't use the NPU in the Ryzen 7040/8040, Phoenix and Hawk Point (XDNA). No NPU, or one it can't use? Your graphics card or processor does the work. What Castellan does with your NPU.