Accuracy-preserving choices
W8, PCQ4, SPQ4, KARQ P1/P4, and automatic mixed precision are implemented across the current DS-CNN validation matrix. Every MicroQuant candidate clears the accepted 90% quality floor.
Technical demo · commercial evaluation funnel
Fit more model
into less silicon.
MicroQuant turns supported neural networks into compact, deterministic firmware for constrained microcontrollers. This site presents the current technical demonstration, exact-board quality evidence, memory footprint, and the pathway to a model-specific commercial pilot.
- Best current top-1
- 92.24%
- Smallest W4 product flash
- 24,566 B
- MicroQuant peak SRAM
- 30,712 B
- Accepted quality cells
- 12 / 12
W8 on the frozen Speech Commands v2 acceptance role.
PCQ4 under the validated generic-SWAR build.
Measured footprint for the validated generic-SWAR cells.
Every W8, W4, and mixed MicroQuant physical cell exceeded 90%.
Current platform
A compact deployment stack, scoped to your product.
The compiler, generated assets, and zero-heap runtime form one deterministic delivery path. Format and precision choices are evaluated against the customer's accuracy, flash, SRAM, graph, and target.
Flash and SRAM visibility
Every physical cell carries complete firmware flash, product-attributable flash, static internal SRAM, activation arena, stack, and peak internal SRAM accounting.
From assessment to pilot
A paid assessment confirms graph coverage and footprint fit. A scoped pilot then delivers generated assets, target integration, acceptance evidence, and licensing terms for the customer's product.
This website is a technical demo and commercial funnel.
It is not a self-serve production compiler and does not distribute customer-ready generated assets. Production delivery is model-, data-, board-, toolchain-, and license-specific.
W8
Highest current KWS accuracy and the original stricter-gate pass.
PCQ4
Per-channel INT4 control with the smallest current W4 product flash.
SPQ4
Per-block INT4 scales for locally varied weights and structured sparsity.
KARQ + mixed
Exponent-palette and layer-selection options for model-specific tradeoffs.
Current exact-board results
Quality and footprint across the complete ABI-v6 KWS matrix.
One frozen 12-class DS-CNN, one untouched 4,769-record Speech Commands v2 acceptance role, and fourteen physical cells under generic-SWAR/PIE on the ESP32-S3-DevKitC-1 N16R8 with ESP-IDF v6.0.2.
Validation complete
14 physical cells; three independent clean sessions per cell.
Quality accepted
12 MicroQuant W8/W4/mixed cells; all above 90% top-1.
Exact scope
Sensorless frozen-feature KWS; no live-audio, energy, or cross-ISA claim.
Speech Commands v2 DS-CNN
Accuracy, flash, and peak internal SRAM from the sealed physical matrix.
| Configuration | Backend | Top-1 | Product flash | Complete firmware flash | Peak internal SRAM |
|---|---|---|---|---|---|
| TFLM control | Portable | 91.99% | 262,944 B | 262,944 B | 55,936 B |
| TFLM + ESP-NN control | ESP-NN | 91.99% | 291,424 B | 291,424 B | 56,400 B |
| W8 | generic-SWAR | 92.24% | 34,678 B | 201,216 B | 30,712 B |
| W8 | PIE | 92.24% | 34,909 B | 201,424 B | 30,744 B |
| PCQ4 | generic-SWAR | 91.36% | 24,566 B | 191,120 B | 30,712 B |
| PCQ4 | PIE | 91.36% | 24,797 B | 191,312 B | 30,744 B |
| SPQ4 g8 | generic-SWAR | 91.70% | 34,118 B | 200,656 B | 30,712 B |
| SPQ4 g8 | PIE | 91.70% | 34,349 B | 200,864 B | 30,744 B |
| KARQ P1 | generic-SWAR | 91.40% | 25,998 B | 192,528 B | 30,712 B |
| KARQ P1 | PIE | 91.40% | 26,229 B | 192,752 B | 30,744 B |
| KARQ P4 | generic-SWAR | 90.14% | 34,818 B | 201,360 B | 30,712 B |
| KARQ P4 | PIE | 90.14% | 35,049 B | 201,568 B | 30,744 B |
| Automatic mixed | generic-SWAR | 91.03% | 30,838 B | 197,392 B | 30,712 B |
| Automatic mixed | PIE | 91.03% | 31,069 B | 197,584 B | 30,744 B |
“Product flash” is the product-attributable linked footprint defined by the sealed comparison contract. “Complete firmware flash” includes the full application image under the same accounting rules. Accuracy is a property of each frozen quantized asset; generic-SWAR and PIE use the same asset and therefore share the same top-1 result.
The accepted commercial quality floor is 90% for W4 and mixed candidates.
The earlier, stricter preregistered host gate is preserved separately;
W8 is the only MicroQuant configuration that passed that original gate.
Download the footprint-only machine-readable result →
Bit-exact execution
Every MicroQuant cell matched the declared host predictions and logits on the physical board.
Repeatable builds
Each cell was compiled and validated across three independent clean sessions with content-addressed artifacts.
Bounded claims
The evidence applies only to the named model, board, SDK, feature role, and frozen acceptance set.
Deployment coverage
What can be evaluated and delivered today.
Standard capabilities, limited subsets, custom engagement paths, and out-of-scope cases are separated so a buyer can qualify fit before a pilot.
Supported standard path
Limited stated subset
Custom scoped engagement
Outside scope different solution required
Operators
| Dense / Gemm / MatMul (constant weights, bias, transB) | Supported |
| ReLU / Clip(clamp) fused as dense/conv activations | Supported |
| Identity / Flatten / Reshape (shape-safe) | Supported |
| Conv2D — 1×1 pointwise & 3×3, stride 1/2, same/valid | Supported |
| Depthwise Conv2D 3×3 (channel multiplier 1) | Supported |
| MaxPool / AvgPool 2×2 s2 · GlobalAveragePool | Supported |
| Softmax (argmax / top-k on int32 logits) | Limited |
| Hidden-layer sigmoid / tanh (fixed-domain LUT) | Supported |
| Attention / transformer, GELU, final-layer sigmoid/tanh | Rejected |
Quantization Backends & Import Formats
| spq4 — signed-int4-per-block-fixed-scale; operators: dense, conv2d, depthwise_conv2d | default · portable-runtime · promoted physical evidence scoped by environment |
| karq — signed-int4-exponent-palette; operators: dense, conv2d, depthwise_conv2d | selectable · portable-runtime · promoted physical evidence scoped by environment |
| pcq4 — per-channel INT4 comparison baseline | Fully implemented · evaluation baseline |
Raw float32 weights .bin / .raw | Supported |
| ONNX — Gemm / MatMul / Relu / Clip / Identity / Flatten / Reshape (strict subset) | Supported |
| cnn-json explicit NHWC spec | Supported |
ONNX Conv → NHWC conversion (static CNN subset) | Supported |
| TFLite ingestion (stdlib flatbuffer reader, documented subset) | Supported |
| PyTorch / Keras (via export to ONNX) | Indirect |
MCU Families & Boards · ISA Features
rp2350-arm-m33 · board pico2-rp2350-arm (Raspberry Pi Pico 2) | Silicon-validated: SPQ4 dense, generic-swar |
rp2350-riscv-hazard3 · board pico2-rp2350-riscv (Pico 2, RISC-V path) | Cross-compile verified |
generic-cortex-m7 (ARMv7E-M · arm-dsp, simd32, FPU) | Compiles |
generic-cortex-m55 (ARMv8.1-M · arm-dsp + Helium/MVE) | Compiles |
esp32-s3 (Xtensa LX7 · board esp32-s3-devkitc1-n16r8) | Silicon-validated: ABI-v6 KWS W8/W4/mixed under generic-SWAR and PIE, ESP-IDF v6.0.2 |
portenta-h7-m7 (STM32H747 M7 only · board arduino-portenta-h7-m7) | Integration-ready board pack |
generic-rv32imac (RISC-V RV32IMAC) | Cross-compile verified |
Board nucleo-h743zi (concrete BSP template) | Available as a custom board engagement |
| ISA caps: arm-dsp · arm-mve · riscv-p · riscv-v · xtensa-simd · custom:* | Declarable; kernels vary |
Kernel Backends · Runtime Ports · Evidence
Kernel generic-swar (portable INT4 MAC, always-safe fallback) | Available |
Kernel arm-dsp-smuad (Cortex-M4/M7 __SMUAD) | Available |
Kernel esp32-s3-pie (Xtensa PIE packed-S4 bank) | Complete-model exact-board validation · bit-exact |
| Kernels Helium/MVE · RISC-V P/V | Available through target-specific kernel work |
Ports: host · baremetal-generic · cortex-m scaffold · RP2350 Arm/RISC-V · STM32H7 M7 · ESP32-S3 · custom MQ_PORT_HEADER | Supported |
| Evidence — host benchmark & bit-exact parity | Included |
Evidence — QEMU semihosting (labeled qemu) | KARQ dense + CNN passed · QEMU 11.0.2 |
| Evidence — physical validation on real silicon (RP2350 SPQ4; ESP32-S3 ABI-v6 KWS matrix) | Delivered ✓ |
| Evidence — model-attributed flash and RAM accounting | Delivered across the current KWS matrix |
| Evidence — ABI-v6 Speech Commands v2 KWS, sensorless frozen-feature inference | Delivered · 14 physical cells · all W8/W4/mixed candidates above 90% top-1 |
| Evidence — cross-ISA KARQ validation | Not claimed — validated on one ISA to date |
Validated Toolchains & SDKs
| Python — compiler host | >=3.12; tested 3.12/3.13/3.14 |
| ONNX — model import | 1.22; evidence 1.22.0 |
| Arm GNU Toolchain — firmware toolchain | 15.3.Rel1; evidence arm-none-eabi-gcc (Arm GNU Toolchain 15.3.Rel1 (Build arm-15.149)) 15.3.1 20260627 |
| Raspberry Pi Pico SDK — vendor SDK | 2.3.0 |
| ArduinoCore-mbed — vendor SDK — Portenta | 4.6.0 |
| Espressif ESP-IDF — vendor SDK — ESP32-S3 | v6.0.2 |
| QEMU — emulation | 11.0.2; evidence QEMU emulator version 11.0.2 |
Commercial pathway
Start with a decision-ready model assessment.
The first engagement establishes whether MicroQuant fits your graph, quality target, flash budget, SRAM budget, target MCU, and production environment before deeper integration work begins.
Assessment deliverables
- Supported-operator and import-path qualification.
- W8, W4, and mixed-precision quality study on representative data.
- Product flash, complete firmware flash, and SRAM footprint.
- Target and toolchain integration scope.
- Commercial pilot proposal with acceptance criteria and licensing.
Pilot deliverables
- Versioned compiler configuration and generated assets.
- Runtime and firmware integration for the agreed target.
- Host and physical-board parity evidence.
- Reproducible footprint report and artifact manifest.
- Handoff documentation for the customer's firmware team.
Constrained embedded ML
Dense networks, compact CNNs, keyword spotting, and anomaly models where accuracy, flash, SRAM, and deterministic allocation matter.
Graphs outside the current surface
Attention-heavy networks, dynamic shapes, unsupported activations, or targets without an agreed integration path require separate scoping.
Commercial licensing
Clear evaluation and production boundaries.
MicroQuant is proprietary commercial software. Evaluation access supports technical review; shipping products require a commercial license and a customer-specific validation record.
Commercial product use
- Per-product or portfolio licensing.
- Source, generated-asset, and integration scope defined in writing.
- Support and maintenance terms matched to the product lifecycle.
Evaluation use
- Internal technical review under the evaluation terms.
- No production, resale, or customer-facing deployment.
- No model or dataset upload through this website.
Start a conversation
Tell us what needs to fit.
Share the model class, target MCU, quality requirement, flash budget, and SRAM budget. We will respond with a clear fit assessment and a scoped next step.
What the first review covers
Model fit
Graph coverage, import path, calibration data, and accuracy target.
Memory fit
Flash, SRAM, static allocation, and integration constraints.
Commercial fit
Assessment scope, pilot deliverables, licensing, and product roadmap.