Palette Neat v0.4.0 Release

Announcements
1

Hello everyone,

We are thrilled to announce that Palette Neat 0.4.0 is officially available! It was released on August 21, 2026! :rocket:

ga-dark_0_4_0
ga-dark_0_4_01920×1080 211 KB

This is one of our largest Neat releases so far. It expands Palette Neat with direct MIPI capture, H.265 video workflows, host-side PCIe inference, much faster model loading, expanded GenAI support, eight new applications, and several key Insight improvements.

:sparkles: What Palette Neat 0.4.0 Brings to the Table

  • Direct MIPI Camera Input: CameraInput now provides libcamera with a DMA-BUF pool backed by Neat-owned SiMaAI memory. Compatible camera frames enter the graph without another CPU copy, and Neat securely manages each buffer until the application releases it.
  • Modalix PCIe Inference: Host applications can now run Neat inference on a Modalix PCIe Card via public C++ and Python APIs. This supports synchronous/asynchronous execution, request-correlated queues, timeouts, backpressure, and queue reuse.
  • H.265 Video Workflows: Neat can now ingest and decode H.265 Main-profile RTSP streams. Applications can run inference on decoded frames while seamlessly forwarding the original H.264 or H.265 stream to Insight without re-encoding, keeping all frame IDs and metadata perfectly correlated.
  • Significantly Faster Model Loading: Model loading times have been drastically reduced—without requiring changes to existing model packs or the public Model API. In recorded tests, YOLO26 loading dropped from 32.47s to 1.63s, and a 150 MB model pack loads roughly 56 times faster (278.65s to ~4.96s).
  • Expanded GenAI Workflows: LLiMa 0.4.0 introduces EAGLE3 speculative decoding, multi-ELF execution for larger vision encoders, and Gemma 4 support. Supported models now utilize up to 8K contexts with optimized memory use. Additionally, tool calls now use a consistent response format across supported Gemma, Llama, Qwen, Qwen-VL, Mistral, and LFM models.
  • Eight New Applications: Apps 0.5.0 introduces comprehensive new examples, including MIPI camera capture, thermal face detection, multi-stream pose estimation, SuperPoint feature extraction, SSD-MobileNet detection, PCIe high-density detection, YOLO26 tiny-drone tracking, and the new local Neat GenAI Studio.
  • Stronger Insight Visualization: Insight v0.0.7 adds H.265 browser playback and native segmentation overlays for masks, contours, labels, and boxes. It also brings persistent test media between sessions, interactive API docs (/api/docs), and expanded diagnostics to easily distinguish source, delivery, and decoder problems.
  • More Repeatable CLI Workflows: sima-cli v2.1.16 adds the sima-cli neat download TARGET command for fetching packages without installing them. It also simplifies SDK entry commands, auto-selects the correct Model Compiler for AMD64/ARM64, and improves PCIe discovery.

:books: Dive Into the Release Notes

Palette Neat 0.4.0 is delivered as a coordinated set of components. Review the component-specific release notes for complete changes, installation guidance, and additional compatibility information:

:warning: Important Compatibility Note

Palette Neat 0.4.0 introduces Neat Library ABI version 4. Customers using native C++ applications or plugins must rebuild them against the v0.4.0 libraries. Please review the release notes relevant to your workflow before upgrading to ensure you meet all matching package-version requirements.

:link: Get Started Today

The SDK installs the complete Neat development environment, including the matching runtime, Core library, Insight, and development tools:

sima-cli neat install sdk@v2.1.3.0

We can’t wait to see what you build! If you have questions or feedback, please drop a comment below.

— The SiMa.ai Neat Team

3 Likes