Structured ML

xLSTM,
native in Eldric.

Structured-ML workloads — policy execution, time-series forecasting, vision-language encoding, associative retrieval — run natively inside Eldric on the xLSTM model family, at predictable latency and memory cost. No separate service to deploy or maintain.

Why xLSTM

The transformer alternative for sequence-heavy work.

Transformers dominate text generation, but their attention cost becomes a bottleneck for long sequences, real-time control loops, and any workload where memory budget matters more than parameter count. The xLSTM family restores linear-time recurrence with modern training — a fit for control policies, multi-horizon forecasts, and dense retrieval at scale.

In Eldric these workloads are built in. They run inside the platform itself and are reachable behind a stable, secure surface, so your people can call them from the chat shell, from agentic flows, or from external integrations — with nothing extra to install and nothing to keep running on the side.

Workloads

Four production surfaces, one engine.

Policy execution

Run a trained xLSTM policy on streaming observations and emit actions in real time — for robotics, industrial control loops and autonomous decision-making. A safety fallback engages automatically if a control step misses its deadline.

Time-series forecasting

Multi-horizon forecasts on telemetry streams — sensor data, demand, load, financial series, OEE metrics. Plan ahead and get early warning, without shipping the data off-site.

Vision-language encoding

Turn images and visual scenes into machine-usable representations for perception and search tasks, frame by frame across image and video streams.

Associative retrieval

High-recall lookup against a learned associative memory — the closest stored match returned in microseconds on CPU alone, no GPU required. Complements the standard vector store for rapid lookup and fuzzy matching.

Operating model

How it runs.

Built in, not bolted on: the four workloads run natively inside Eldric — there is no separate service to deploy, supervise or keep healthy. A plain install already serves them.
One secure surface: requests reach the workloads through the same authenticated, TLS-terminated front door every request takes, and the platform dispatches each one to the right workload for you.
Your models, your hardware: load and replace the models these workloads use through the standard model-management interface. Everything stays on machines you control.
Hardware fit: a GPU is recommended for the heavier workloads, with CPU fallback for lighter ones — associative retrieval runs in microseconds on CPU alone. Multi-GPU scaling is supported where it helps.
Licensing: the four workloads are available now; per-tenant quotas and advanced features scale with your license tier, opted in from the admin console.

When to use which workload

Decision shape.

Predict the next N steps of a sensor series? Time-series forecasting.
Pick the next action given the current observation? Policy execution.
Compress images or video frames into representations for downstream search? Vision-language encoding.
Look up the closest stored match for something seen before? Associative retrieval (use the standard vector store if you need semantic search instead).

Status

Where it is on the roadmap.

All four workloads are live in 5.0 and ship inside the standard Eldric install (and the GPU-enabled variant) — nothing separate to add. Per-tenant limits and advanced features such as custom-architecture training and multi-tenant policy isolation scale with your license tier.

Coming in 5.0.x: additional model architectures in the policy and encoder slots, and deeper integration with the media and industrial-sensing capabilities.

For developers — full API reference + integration patterns
Features — how structured-ML fits alongside the rest of the platform
Robotics — control policies and real-time decision loops in the field
Help & documentation — the master index

xLSTM,native in Eldric.