Q4NX conversions become usable, with explicit limits

The release says converted Q4NX models now load from the NPU model list in text-only mode. Compatibility labels distinguish models that run directly, need conversion, convert but will not run, or remain unsupported. The converter also rejects GGUF quantizations known to produce unusable output.

There is a documentation inconsistency worth noticing: the NPU guide’s FAQ still says loading custom conversions is not yet supported. For this specific change, the dated beta release notes are more recent evidence. Users should retain the exact application version when reporting a conversion problem.

Hardware and runtime requirements still apply

The installation guide specifies an AMD XDNA 2 or newer NPU, supported Ryzen AI families and Windows 11 or Linux x64. Linux requires the appropriate amdxdna driver, firmware and memory-lock configuration. An ordinary AMD processor without the supported NPU is not equivalent hardware.

Q4NX is a runtime-specific model format, not a guarantee that every GGUF can be loaded unchanged. The guide describes conversion on the CPU and inference through FastFlowLM. These steps are separate from training an adapter or changing a model’s task behavior.

API access does not remove memory constraints

AMD describes compatible agent applications reaching local inference through Unsloth Desktop’s OpenAI-compatible API, with Lemonade and FastFlowLM handling the execution path. The same vendor example says a full-context Qwen3.6 35B A3B configuration needs around 28 GB available to the NPU, corresponding to 64 GB system RAM under the stated Windows sharing assumption.

That is an example configuration rather than a universal minimum for every model. Context length and the selected weights matter. We have not measured speed, battery life or energy use, so we do not transfer advertised efficiency ratios to an arbitrary laptop.

LoRA training is another part of the update

The release adds LTX-2.5 and MOSS-Image LoRA training support. This should not be read as a claim that the NPU trains those adapters: the documentation assigns supported training and image or video generation workloads to the Radeon GPU. Local inference and fine-tuning solve different problems.

For a task-specific assistant, our suggested acceptance test combines model correctness with the surrounding application: compare representative answers, tool requests and failures before and after changing the runtime. Check the selected weights’ license separately. A software integration does not replace that license or establish that the assistant is authorized to execute every operation it can describe.