Appearance

HLM-Micro — Edge & MCU Variant ​

HLM-Micro is Qriton's polynomial-Hopfield variant designed from the ground up for microcontroller-class hardware: ESP32-S3, Alif Ensemble, Kendryte K230, BL808. Not a quantisation of HLM3 — a separate architecture that inherits the same multi-basin energy landscape but is purpose-built for sub-MB flash, sub-watt deployments.

Access

Trained weights and deployment artefacts are available through the Early Access program or a commercial engagement. This page describes what the architecture is and where it deploys.

Why a separate variant ​

HLM3 doesn't fit on an MCU, and simply quantising it down doesn't solve the problem. HLM-Micro inherits the polynomial d=3 Hopfield dynamics, drops the architectural pieces that don't earn their parameter cost at this scale, and adds features specific to the edge envelope:

  • Runtime compute dial — the same weights serve a low-power emergency mode, a default deploy mode, and a quality mode. User selects at inference time, no retraining.
  • Basin paging — hot-set basins resident in fast memory, warm in PSRAM, cold paged from QSPI flash. LRU eviction with async prefetch.
  • Multimodal single checkpoint — a short modality prefix lets one model handle text, audio, sensor, and vision-crop inputs through a shared 1D-conv stem.
  • Hash-chain audit trail — every inference emits a SHA-256 digest binding input, weights, and basin-ID trajectory. Tamper-evident, replay-verifiable.

Architecture at a glance ​

PropertyValue
Polynomial degree3 (shared with HLM3)
Param range1 – 5M
FFNnone — pure Hopfield blocks
Precision targetINT4 / INT8 stratified quantisation
Flash footprint targetsub-MB
Active-power targethundreds of mW
T dialthree operating modes

Exact architectural dimensions, precision stratification, and basin-paging details are shared with Early Access partners under the standard evaluation terms.

Status ​

v0 MVP validated. Multiple trained checkpoints across distinct tasks — industrial anomaly, gesture recognition, human activity, ECG anomaly. One architecture, different heads per task. See the Model Zoo for the published checkpoint set.

Headline result ​

On UCI HAR (30 subjects, subject-disjoint test split, real human IMU data), HLM-Micro reaches 89.75% test accuracy — in the published TinyML baseline range of 85–95% on the same benchmark. Full model card: hlm-micro-har-v0.

Deployment shape ​

Two supported patterns:

  1. Host-inference — MCU reads sensors and streams feature frames to a laptop / gateway that runs HLM-Micro and emits audit certificates. Simpler integration, full certificate support, good for pilots.
  2. On-MCU inference — model runs natively on ESP32-S3-class hardware. Custom C runtime with INT8 deployment. Partnership-level engagement.

Energy Language integration ​

HLM-Micro models use the same BasinSurgeon API as HLM3 / HLM-Spatial / HLM-Audio. The basin centroids are addressable; the Energy Language operations apply identically, just at smaller scale.

Intended workflow:

  1. Capture / edit concepts on an HLM-Micro model using Energy Language on a laptop
  2. Export the modified basin as a small OTA patch
  3. Push the patch to deployed MCUs over standard update channels
  4. Devices reload the affected basin without a full firmware reflash — concept edited in the field

Commercial positioning ​

HLM-Micro fits best where standard TinyML doesn't:

  • Regulated industrial sensors where audit-trail requirements rule out standard TinyML
  • Defense / critical-infrastructure edge nodes requiring EU-sovereign stack
  • Multimodal smart-building devices where one MCU handles audio + vibration + commands on one model
  • Medical wearables where regulator-facing explainability is part of the product

Not positioned for: consumer smart speakers, hobbyist IoT, or any environment where a general-purpose SBC is acceptable.

Contact us for partnership or pilot-project discussion.