Voyager® Software Development Kit (SDK) v1.9 is here, and it puts more of the platform in your hands than any release before it. If you read our announcement of AxScript in September, this is the release where AxScript models start to ship. If you followed the Pipeline Builder debut in v1.6, this is the release where it reaches Beta, with faster pipelines, native Windows support, and a model zoo that now arrives as standalone Python recipes.
Release Highlights for version 1.9
Top 3 highlights:
- AxScript as a new programming language for Axelera AI platforms, along with new LLMs, VLMs, and ViTs, delivered as AxScript models
- Build faster pipelines with the upgraded Pipeline Builder, now native on Windows
- Standalone Python recipes for models in our Model Zoo
Also new in this release:
- AxGraph, our new experimental MLIR-based graph compiler for generative AI and transformer models
- A portable .axe format that records a pipeline end to end
- New task types: RetinaFace face detection, YOLACT instance segmentation, and YOLO26 monocular depth
- Preprocessing kernel fusion in Pipeline Builder API & performance improvements
- Vulkan-based rendering, now the default renderer, and tracker support on Windows
- Multi-stream scheduling, GIL-free multi-threaded scaling, and a new per-model profiler in the Pipeline Builder
- A new Executor V2 (Beta) in the runtime, with synchronous and asynchronous execution modes
- New product names: Axelera Embedded 110m / 113m / 111c, Axelera Edge 130p / 232p, and Axelera Server 150p replace the old Metis board names
AxScript models: generative AI and transformers on Axelera AIPUs
AxScript, the Python-based domain-specific language we introduced in AxeleraScript: One Language, every Inference workload solved, gives developers the capability to program our silicon at a low level, offering high programmability, control and flexibility. Lead customers have already created a lot of value by implementing models in AxScript; Voyager SDK 1.9 is where you start to benefit from it too!
AxScript models ship as code, so you can read exactly how each one is implemented and how it runs in our programming models. You can also re-use the code, make modifications and adapt it to your own models based on your needs. You will find the code under model_axscript/, with one Python package per model which includes all the code and kernels and the AxScript language specification and guide in our documentation.
This release offers a range of generative AI and transformer-based models, including LLMs, VLMs, and ViTs, with support for several state-of-the-art models in each family.
| LLMs | qwen3-0.6b |
| qwen3-1.7b | |
| qwen3-4b | |
| qwen3-8b | |
| llama-3.2-1b | |
| llama-3.2-3b | |
| llama-3.1-8b | |
| lfm2-700m | |
| lfm2.5-350m | |
| lfm2.5-1.2b | |
| domyn-10b | |
| VLMs | qwen3-vl-2b-instruct |
| qwen3-vl-8b-instruct | |
| smolvlm2-256m | |
| smolvlm2-500m | |
| smolvlm2-2.2b | |
| lfm2.5-vl-450m | |
| lfm2.5-vl-3b |
Additional information can be found in the release notes.
Pipeline Builder: faster, more capable, and on Windows
The Pipeline Builder API introduced in v1.6 lets you define an entire inference pipeline as composable Python, and to build and execute the pipeline from the same Python programming unit. In v1.9, the Pipeline Builder API reaches Beta and is enhanced with new features and performance improvements. For example, on Yolo-based detection models, pipelines built using the Pipeline Builder API can now achieve the same performance as the legacy, YAML-based pipelines for medium (m), large (l) and extra large (x) variants of the models, while further improvements in upcoming releases will achieve parity also for small (s) and nano (n) variants.
The Pipeline Builder API makes for very compact, readable and maintainable code and new features such as the capability for the user to bring their own postamble kernels add a lot of flexibility. As an example, a pipeline for Yolov8 using a custom kernel implementation for the postamble can be constructed as follows:
| POSTAMBLE = Postamble(fused_kernel='UltralyticsYoloPostamble', num_classes=80, dfl_bins=16) pipeline = op.seq( op.color_convert(dst='RGB'), op.letterbox(width=640, height=640), op.to_tensor(), # axm_path is the path to the compiled model op.load(str(axm_path), using_postamble=POSTAMBLE, device_config=DEVICE_CONFIG), op.decode_detections(algo='yolov8', num_classes=80), op.nms(), op.to_image_space(), op.ax_detection(class_id_type=op.CocoClasses), ) for _frame, results in pipeline.stream(video, max_in_flight=1): # show detections |
Notable new features and enhancements include:
- A new fusion pass runs pre-processing as a single fused OpenCL kernel, up to 10x faster than OpenCV.
- New scheduler defaults achieve much higher throughput (e.g., +81% for Yolov8n).
- The portable .axe format records a pipeline end to end, so you can package it and hand it off with confidence.
- Kernel operators scale across threads, no longer serializing on Python's GIL: an 8-thread pre-processing workload went from 0.2x to 7.8x scaling, and zero-copy buffer import makes 1080p and 4K resizes to 640x640 3x faster.
- A new profiler attributes each operator to its model and its role, separates work from waiting, and reports device occupancy and in-flight depth.
- New task types and operators include RetinaFace face detection, YOLACT instance segmentation, a fast NMS algorithm, decoders for every shipped head, and perspective and barrel-distortion correction.
- Live sources keep the freshest frame instead of falling behind, files can play back at a chosen rate, Vulkan is the default renderer, and a low-latency preview can run ahead of inference.
Native Windows support. Pipeline Builder now runs natively on Windows through an MSVC win_amd64 wheel of axelera-runtime2, for Python 3.10 to 3.14. It includes a Vulkan-based renderer, and trackers are supported. Compile on Linux or WSL2, then run your pipelines natively on Windows.
Model Zoo recipes: one Python file per model
The zoo now reaches you as standalone Python recipes in model_recipes/. The Python zoo (axzoo), introduced as an Alpha in v1.8, is now Beta and the primary way models are built, run, and evaluated. The YAML-based zoo still ships in v1.9 and remains supported, but from the next release new models will be added to axzoo only.
Each recipe is a single, self-contained file that shows the whole path a production deployment takes:
- Calibration
- Quantization
- Compilation with the model's tuned settings
- The end-to-end inference pipeline, wired in a few lines with Pipeline Builder
Every setting that shaped our published numbers is a line you can read and change.
Run it, measure it, reproduce it. Run a recipe to get throughput and a rendered result window right away. Add --eval (install with pip install axelera-zoo[eval]) to reproduce the published accuracy on the same scoring path we used for the figures.
Three commands cover the workflow:
- axzoo list shows the catalog.
- axzoo info <model> shows where a model comes from upstream and how it is exported.
- axzoo fork <model> copies the recipe, its config, or its export script into your own project, so a reference model becomes the starting point for your own weights.
The full list of recipes is in the release notes.
For the full release notes, documentation, and technical support, visit the Axelera AI Customer Portal.
Try it out
Installation via pip is now the recommended path. As in v1.8, the full SDK installation requires both axelera-devkit[all] and axelera-rt packages:
| # Clone out the repository and install dependencies # Install the packages |
Build with Axelera Wingman
You can put everything in this release to work with Axelera Wingman, our agentic partner for the Voyager SDK. It is built with the Voyager SDK knowledge base and stays current with the latest software updates and documentation, so version v1.9 is covered from day one.
Describe what you want to build in plain language, and Axelera Wingman turns it into working AI: it builds the Axelera pipeline, runs it on real hardware, and iterates with you. Whether you want to explore the new model recipes, set up a Pipeline Builder pipeline, or find your way around the new AxScript models, Axelera Wingman knows where to start.
Launch Axelera Wingman and test it for yourself. It is free, and you can register for an account through the Community.
We'd love to hear what you build with it! Comment below or share your project in Axelera's Community.
