hivyr.io
Neural audio inference IP

State space,
shipped.

A Mamba-3 SSM inference engine in dependency-free C++17 — running as CLAP and VST3 in Ableton Live and Bitwig today. No ONNX Runtime, no GPU requirement, no research-program overhead for your team.

FIG 01 — MAMBA-3 SELECTIVE SCAN continuous signal → discrete state steps
Verified, not claimed

Every number traces to a CI artifact.

The engine's test suite is the marketing page: clap-validator passes, a three-OS build matrix, and bit-exact Python↔C++ parity run on every commit.

hivyr — engine --verify
engine mamba-3 SSM · MIMO R=4 · time-domain
parameters ~1.66M trainable · 6 layers · d_model=128 · d_state=32
latency 0 samples reported (time-domain engine)
audio thread 0 heap allocations — RT-safe by construction
parity bit-exact python ↔ c++ (CTest)
formats CLAP ✓ · VST3 ✓ (clap-wrapper) · ARA 2.0 (scaffolding)
dependencies C++17 stdlib only — no ONNX, no GPU runtime
weights single flat binary model.bin · mmap · version-checked
The package

Four components. One drop-in engine.

◆ Mamba3Core

The C++17 inference engine as a static library. Per-sample, time-domain processing with stack-allocated state — zero heap allocations on the audio thread, deferred atomic resets, and a lock-free SPSC queue for UI communication. Link it like any other library; it has no runtime dependencies.

◆ Plugin shells

Production CLAP implementation, VST3 via clap-wrapper (verified in Ableton Live), and ARA 2.0 extension scaffolding. Reference hosts: Bitwig, Carla, REAPER.

◆ Adaptation pipeline

Layered per-effect training against your DSP targets — saturator, compressor, denoiser profiles from a shared frozen base. Delivered as versioned model.bin weight files.

◆ Documentation & SLA

Integration guide, weight-format specs (V2 shipping / V3 roadmap), and tiered support with response-time commitments.

◆ Source escrow (Tier F)

Full-IP licenses include escrow with release on insolvency or abandonment — your product line is never hostage to ours.

Roadmap, dated

What ships, what ships next.

SHIPPING
R=4 time-domain engine live 0-sample reported latency · CLAP/VST3 · bit-exact parity · 3-OS CI
IN PROGRESS
R=8 spectral engine training validated STFT frontend (n_fft=2048 → 1024-sample latency), complex-pole SSM, CREPE f0 conditioning, cross-track attention — C++ rewrite in progress
NEXT
Per-effect model catalog Saturator / Compressor / Denoiser variants from the shared frozen base
Honesty policy: features on this page are labeled shipping, in-progress, or roadmap. The 0-sample latency figure applies to the current time-domain engine; spectral-engine figures will be published with measurements, not estimates.
Evaluation

Run it on your bench in 14 days.

A watermark-limited evaluation plugin against one of your reference targets. No fee, no obligations — just measurements.

Request evaluation → Read the technology page