Palette SDK 3.0 is Here! πŸš€

The day is finally here! :rocket:

We are thrilled to announce that Palette SDK 3.0 is officially available!

Released on October 31, 2026, this marks the biggest and most impactful software release in SiMa.ai’s history. Bringing together work from across all our teams, it dramatically expands what you can run on Modalix and makes complete applications easier to build, measure, and deploy.

We are also incredibly excited to share that the Neat library has been upgraded to v0.6.0! The core C++ and Python Model, Graph, and Run APIs are preserved. Existing customers can seamlessly take advantage of the new capabilities by recompiling native applications against the Palette 3.0 SDK, installing the matching packages, and validating their workloads.

:sparkles: Palette 3.0 Release Highlights

1. New GenAI capabilities on Modalix and over PCIe

Bring local reasoning, visual understanding, and voice interfaces into existing products.

  • Native PCIe support for LLMs, VLMs, and speech recognition for host-based products.
  • New Mixture-of-Experts (MoE) support, opening a new class of edge models (GPT-OSS-20B has been confirmed running on a 16GB Modalix system!).
  • Full Qwen 3.5 support (0.8B to 9B) and concurrent language/vision processing.
  • Faster generation with up to 3.5Γ— higher decoding throughput using new speculative-decoding options like DFlash and Gemma 4 multi-token prediction.

2. More applications on one system & simpler deployment

  • Container Consolidation: Independent processes and Docker containers can now share the accelerator and video engines. Use standard Docker workflows to build, publish, and deploy on the DevKit. The new Docker example runs six apps side by side.
  • New Execution Foundation: Kernels, Runner, Scheduler, and Codecs now operate as dedicated components with direct runtime/driver integration for better memory sharing.
  • Custom Kernels: Custom EV74 and A65 kernels can now be used directly in Neat. View Custom Kernel Docs.

3. More cameras, faster detection, and stronger drone tracking

  • Higher Density & Faster Processing: Camera-stream profiles have increased from 48 to 60 (a 25% increase). RF-DETR Small throughput has advanced to roughly 120 FPS on 1080p video.
  • Tracking & Codec Upgrades: Enjoy improved TrackKLT latency/correctness for drones, native USB webcam support, and more capable hardware codecs (H.265 decoding up to ~730 FPS aggregate at 720p).

4. Measure real workloads and see the complete system in Insight

  • Deep Profiling: The improved Benchmark API and TelemetryMonitor let you measure p50/p95/p99 latency, per-layer statistics, and system power over specific intervals.
  • Unified Insight Interface: Features a new Sentinel-backed Stats page, a dedicated GenAI Studio tab (chat, vision, speech), and a Peripherals tab for easy hardware verification.

5. A stronger platform for development and field deployment

  • Hardware & OS: Broader integration (Raspberry Pi host-driver, expanded IMX568 modes), Debian 13/eLxr, Docker support, and faster boot times with lower idle consumption.
  • Maintenance: Easy A/B upgrades, USB-drive recovery, automatic NVMe mounting, and offline deployment bundles to drastically reduce setup and maintenance friction.

:warning: Before Upgrading

Neat 0.6.0 requires Platform 3.0.0 on the DevKit.

Rebuild native apps against the 0.6.0 headers (ABI 6) and reinstall PyNeat. If you chain models across separate graphs into an EV74 stage, call .cvu() on the output first. RTSP inputs now drop late packets by default.

Moving from 2.1.3 to 3.0.0

A regular update cannot convert a 2.1.3 board because 3.0.0 moves to Debian 13 with A/B images.

Please carefully follow the full migration instructions:

Convert from 2.1.3 to 3.0.0

:books: Dive Into the Release Notes

Palette SDK 3.0 is delivered as a coordinated set of components. Review the component-specific release notes for complete changes and compatibility information:

:link: Get Started Today

Install the Neat SDK

The SDK installs the complete Neat development environment, including the matching runtime, Neat Core, Insight, and Model Compiler 3.0.0.

sima-cli neat install sdk@v3.0.0.0

Optional: Update only Neat Core

sima-cli neat install core@v0.6.0

Optional: Install Neat Apps

sima-cli neat install apps@v0.6.0

By default, the Apps installer installs its compatible Neat Core and the latest Insight. If the SDK already provides them, skip both:

NEAT_APPS_SKIP_DEPENDENCIES=1 sima-cli neat install apps@v0.6.0

This release was a massive cross-team effort. Thanks and congratulations to everyone who made it happen!

We can’t wait to see what you build with Palette 3.0! :rocket:

If you have questions or feedback, please drop a comment below.

β€” The SiMa.ai Team