Most of the AI chatter this year is about giant models in data centers. Yet some of the most practical progress is happening at the opposite end: tiny chips in rings, earbuds and sensors that must run a neural network for days on a sliver of battery. On November 9, 2026, Ambiq Micro announced heliaAOT 1.0, a compiler built for exactly that job. According to Ambiq’s press release, it cuts inference latency by up to 76% and energy per inference by up to 72% against LiteRT Micro, a common embedded runtime.
What heliaAOT actually does
heliaAOT is an ahead-of-time compiler for the company’s HELIA AI ecosystem. Instead of shipping a general-purpose interpreter that reads a trained model while the device runs, it converts the model into optimized C code before deployment. That means less processing time, less energy and less memory spent every time the device makes a prediction.
The approach matters because memory-constrained devices cannot afford much overhead. An interpreter has to carry generic machinery for every operation it might meet. A compiled model carries only what that specific network needs. Ambiq says heliaAOT has already been deployed in mass-produced products, and that it is in discussions with some larger customers about adopting it in current products.
The numbers Ambiq reports
Ambiq published benchmark results measured on its Apollo510 chip at 96 MHz, compared with LiteRT Micro under matched per-model conditions. The gains differ by workload:
- Health monitoring (ECG, PPG and sleep models): 76% lower latency, 72% lower energy per inference and 56% smaller firmware images.
- Vision: 61% lower latency, 51% lower energy and 29% smaller images.
- Voice and audio: 52% lower latency, 45% lower energy and 47% smaller images.
- MLPerf Tiny reference models: 36% lower latency, 27% lower energy and 37% smaller images.
The company adds useful caveats. Energy is gross energy on the microcontroller’s power rail per inference, so the results are model-level figures, not battery-life measurements. The MLPerf Tiny row uses the reference models and is not an official MLPerf submission. Firmware size means flashed storage, not total peak RAM. These are vendor benchmarks, and developers can reproduce them on their own models using Ambiq’s heliaPROFILER tool.
Broader model support and easier reuse
Beyond speed, the 1.0 release widens what a small device can run. Ambiq says developers can now run supported FP16 and FP32 models without converting them to integers first, a step that often costs accuracy and engineering time. In a matched FP32 comparison across six model families, it reports roughly four times faster inference than LiteRT Micro.
Operator coverage reached 65, plus support for custom operators, so more model architectures can be compiled. Streaming layers such as LSTM, GRU and SVDF are supported too, which suits always-on tasks like continuous voice, biosignal and sensor processing. Median firmware images come out 29% to 56% smaller than with LiteRT Micro, freeing storage and RAM for other features.
There is also a workflow benefit. A supported model can be retargeted across 15 Apollo systems-on-chip with a one-line compiler configuration change, so work done for one product can carry into the next. Over the past 12 months Ambiq delivered 17 heliaAOT updates, expanding operator support from 35 to 65 and Apollo targets from 8 to 15. The tool was also recognized in the Best Tools category at Embedded World 2026.
What it means
For wearable and sensor makers, power and memory budgets decide which AI features are possible at all. A compiler that trims energy and firmware size can turn a feature that needed a bigger battery or a pricier chip into one that fits. Health-monitoring models showed the largest gains in Ambiq’s tests, which is notable for devices that watch signals around the clock.
It also shows where competition in edge AI is heading. Silicon vendors are no longer selling chips alone; they are bundling compilers, profilers and runtimes so customers can squeeze more from the same hardware. For developers, the practical step is to profile their own models before trusting any headline percentage, since results depend heavily on the network and configuration.
Ambiq says heliaAOT 1.0 is available now, with benchmarks and evaluation requests handled through its website. This article reports a company announcement and is not a product endorsement.
The bottom line
Efficient AI on tiny devices is mostly a software story. By compiling models ahead of time, Ambiq claims large cuts in latency, energy and firmware size on its Apollo chips, with the biggest wins in health-signal workloads. The figures come from the vendor’s own tests, but the direction is clear: smarter always-on devices will depend on tools like this as much as on new silicon.