Release v0.6.1

Apps 0.6.1 updates the packaged Neat examples, dependency installer, and GenAI Studio. It improves live demos and PCIe detection handling while keeping the examples aligned with the current Neat Core APIs.

Installation

The Apps installer compares the selected Neat Library version with the installed SDK and paired DevKit versions. It skips matching installations and shows where an installation or replacement is needed before making changes. If the SDK is unpaired or the DevKit cannot be reached, it warns and can install the library in the SDK only.

Insight is also kept when current. When a newer Insight is available, the installer offers an update from main. Declining keeps the installed version. A failed update or service restart prints a warning and allows Apps installation to finish. On a DevKit, an already-running Insight process needs a manual restart to use the updated code. See #677 and #695.

GenAI Studio

GenAI Studio now allows two resident chat models by default. The configured limit can be 1 to 4, and running setup again preserves an existing valid setting unless an environment override is supplied. Invalid non-integer or below-one limits are rejected.

The limit is not a memory reservation: two models may still exceed available accelerator memory. At the resident limit, Insight asks which model to unload; an out-of-memory error can require a backend restart. See #685.

Live examples and configuration

• SuperPoint now uses Insight RTSP with hardware decoding and source-paced playback. Existing local-file configuration using io.input must be updated to source.rtsp_url, with optional source.tcp and source.latency_ms. The packaged README describes setup.
• PatchCore renders RTSP heatmaps directly into NV12 frames and smooths heatmaps at model resolution. Offline BGR processing remains available.
• FastFlow plays its demo at 30 FPS and holds each disc for one second so anomalies are easier to inspect.
• Multi-stream multi-model, BlazePose3D, and drone tracking use Core's automatic input-memory selection instead of redundant overrides.
• Examples remove unnecessary timeouts on blocking calls while retaining short polling loops and application stall deadlines where needed. This requires the matching Core idle/resume fix. See #693 and #698.

RF-DETR examples and application-builder guidance also clarify the default tensor preprocessing behavior. See #688.

PCIe detection

The PCIe high-density detector now reads the negotiated logical BBOX region, including nonzero offsets, instead of treating the start of the raw allocation as detection data. It tracks progress per stream and validates card IDs, queue sizes, buffer capacities, and request timeouts before startup. Legacy BBOX byte streams remain supported. See #690.

Compatibility and validation

Use the Core version and dependency set selected by the Apps package. Core selection follows the configured snapshot policy, so the resolved package identity matters in addition to the Apps source version. Follow each installed example's README and configuration when upgrading.

Individual changes include host, installed-package, and selected DevKit evidence in the linked pull requests. The release branch CI was still running when this draft was prepared; the new tag build must complete before its result can be reported. These checks do not establish long-duration stability, performance, or accuracy for every model and stream configuration.

Full change history from 0.6.0 to 0.6.1