Windows ARM64 NPU activation fails: lic.nexa.ai:443 unreachable during license validation

[Cutcall-Git]

Pre-checks

• I searched existing issues
• I’m using the latest NexaSDK release

What happened?

I’m running Nexa SDK / Nexa CLI on Windows ARM64 with a Snapdragon X Elite device.

The local NPU runtime loads far enough to reach license activation, but activation fails because lic.nexa.ai:443 is unreachable from multiple networks.

The failure appears to happen during first-time online activation/validation, not during Gradio startup, Python app launch, or model-name resolution.

The local runtime appears to attempt:

https://lic.nexa.ai/v1/licenses/actions/validate-key

Observed runtime error:

License validation failed: fail to request license host for activation
NexaLicenseInvalidError: [-100601] Invalid license

Based on local debugging:
- the runtime token shape appears to be accepted
- the model name qwen3-4b is accepted by the NPU plugin
- the runtime progresses into activation/validation
- the next blocker appears to be the activation request to the license host

Steps to reproduce

1. Install Nexa SDK / Nexa CLI on Windows ARM64 / Snapdragon X Elite.
2. Configure a runtime token in NEXA_TOKEN using the key/... token format.
3. Attempt to run local NPU inference using:

qwen3_4b-genie-w4a16-qualcomm_snapdragon_x_elite

4. Backend/runtime path:

plugin_id="npu"
device_id="npu"
model_name="qwen3-4b"

5. Runtime reaches license activation and fails with the license host request error.

Logs & stack traces

License validation failed: fail to request license host for activation
NexaLicenseInvalidError: [-100601] Invalid license

Network diagnostics
Powershell:
nslookup lic.nexa.ai

Result
Name:    lic.nexa.ai
Address: 54.200.56.5

Powershell:
curl.exe -4 -v https://lic.nexa.ai
Result:
ComputerName     : lic.nexa.ai
RemoteAddress    : 54.200.56.5
RemotePort       : 443
TcpTestSucceeded : False

Powershell:
curl.exe -4 -v https://lic.nexa.ai

Result:
* Host lic.nexa.ai:443 was resolved.
* IPv4: 54.200.56.5
*   Trying 54.200.56.5:443...
* connect to 54.200.56.5 port 443 failed: Timed out
* Failed to connect to lic.nexa.ai port 443: Could not connect to server
curl: (28) Failed to connect to lic.nexa.ai port 443


This was tested from both Comcast and a Google Fi hotspot, with the same broad failure mode.

Model(s) and quantization

qwen3_4b-genie-w4a16

NexaSDK version

Nexa CLI: v0.2.73 Nexa CLI Bridge: v1.0.45-rc1

Install method

Nexa CLI / Nexa SDK on Windows ARM64

OS and version

Windows ARM64 Python 3.11 ARM64 Snapdragon X Elite device

Hardware / accelerator

Snapdragon X Elite NPU Backend: plugin_id="npu" device_id="npu"

Additional context

Questions:

1. Is https://lic.nexa.ai currently the correct activation host?
2. Is the host deprecated, moved, or temporarily down?
3. Is there an offline activation path or licenses.txt cache workflow?
4. Should the public key/... runtime token validate offline, or does first activation always require the license host?
5. Is the key/... value from the dashboard the correct runtime/NPU license key, or is there a separate NPU deployment key?
6. Are there alternate endpoints or firewall allowlist entries?

I can provide additional traceback snippets or runtime details if helpful.

[iwr-redmond]

Nexa was acquired by Qualcomm in March and infrastructure is currently being migrated. Have you tried the default license in the Readme file while we wait for the migration to complete?

[Lando-00]

Hitting the same issue here on Surface Pro 11 (Snapdragon X Elite, Win 11 ARM64) - Nexa CLI v0.2.73

Tried the README default license both ways:

• $env:NEXA_TOKEN = 'key/...'
• nexa config set license 'key/...' (then NEXA_TOKEN unset, verified via nexa config list)

Both end up at the same 21.5 s timeout with License validation failed: fail to request license host for
activation. So the persistent-config path doesn't bypass the dead host either.

[iwr-redmond]

You may wish to try the new XLLamaCPP build for ARM64. While it does not work with the proprietary Nexa format it will run any LlamaCPP compatible GGUF file with Vulkan NPU acceleration.

pip install xllamacpp --index-url https://xorbitsai.github.io/xllamacpp/whl/vulkan

[Cutcall-Git]

Nexa

Yes, I did try the default license - referenced it as public key in the issue. That key does open the gate to accessing the models. But the failure is, like Lando said, persistent timeouts failing to contact nexa.ai to validate only once the model is loaded the first time and the user prompts the model.

From what I gather, in one form or another they are re-stablishing NEXA SDK support for Qualcomm Arm64 PCS? So right now, it's mid-transfer?

[Cutcall-Git]

You may wish to try the new XLLamaCPP build for ARM64. While it does not work with the proprietary Nexa format it will run any LlamaCPP compatible GGUF file with Vulkan NPU acceleration.

pip install xllamacpp --index-url https://xorbitsai.github.io/xllamacpp/whl/vulkan

OK I will look into that. TY! I have other options to run local AI. This particular project is really looking to bench the Surface NPU vs the other PCs in the mix. So in the interim, using Vulkan might be good to try.

That uses the NPU and not Adreno?

[iwr-redmond]

So right now, it's mid-transfer?

Yes, I reckon they’re doing a land-office business behind the scenes.

That uses the NPU and not Adreno?

Vulkan works with almost any GPU type, and this should include Adreno.

[fleps]

So Qualcomm bought them, and many months later all seems abandoned? What do we do to run local LLMs with NPU support? The Vulkan suggestion from above is not compatible with Snapdragon NPU's afaik.

[iwr-redmond]

The SDK is being updated in preparation for what I reckon is a relaunch with a new name. In the meantime, Vulkan is supported on Snapdragon.

[fleps]

iwr-redmond yes, I know, but this issue is not about using the Snapdragon GPU, we are talking about the NPU. They are different things.
There's no real benefit on running a LLM in the Adreno GPU, it will be slower than just running using the Snapdragon CPU, specially the X and X2 Elite ones.
Nexa SDK was pretty much the only reliable way and it was progressing ok and now it's completely halted.

[iwr-redmond]

Qualcomm will catch up eventually.

[pinoepel]

Confirming this on a second device and network — same root cause.

Environment: Lenovo Snapdragon X Elite (X1E78100), Windows 11 ARM64 (build 26200), nexa-cli v0.2.73.

lic.nexa.ai:443 is unreachable here too, while general connectivity is fine:

Test-NetConnection lic.nexa.ai    -Port 443  → TcpTestSucceeded: False  (DNS 54.200.56.5)
Test-NetConnection huggingface.co -Port 443  → TcpTestSucceeded: True

Notably, everything except license validation works:

• nexa pull NexaAI/LFM2.5-1.2B-npu downloads and registers fine (nexa list shows it).
• The model loads onto the NPU (nexa infer reaches loading model...).
• It then fails at the call to https://lic.nexa.ai/v1/licenses/actions/validate-key → SDKError(Invalid license).

Two follow-ups for the maintainers:

1. The docs (docs.nexa.ai/en/nexa-sdk-go/npu) still state "previous token validation service has been deprecated — simply set the access token, no validation needed", which contradicts v0.2.73 hard-depending on lic.nexa.ai. Please reconcile.
2. Since sdk.nexa.ai now redirects to the Qualcomm AI Hub "coming soon" page, where should users obtain a valid token, and is an offline-activation path planned while the endpoint is down?

[iwr-redmond]

A new version is out: https://github.nexa.ai/en/run/cli/install/

[AGrunewald]

Has there been any announcement around GenieX ? https://github.nexa.ai/en/get-started/what-is-geniex/ Looks like they decided to no longer support the Snapdragon X Plus? That chip is a year old.