Eldric's recognition + reasoning capability comes in three tiers — same architectural primitive at every scale, just more memory, more experts, more depth as you add hardware. Nexus bundled with every install runs on a Raspberry Pi 5. Nexus Pro takes a workstation. Eldric Cortex takes a datacentre. You pick the tier; the platform handles the rest.
| Tier | What it does | Hardware | Tier | Status |
|---|---|---|---|---|
| Nexus | Recognises themes from every conversation. Microsecond recall. | Any CPU — Raspberry Pi 5 and up | Bundled, all tiers | Available today |
| Nexus Pro | Recognises 100× more themes. Multi-exemplar matching for nuanced domains. | Workstation: 1 RTX 6000 or equivalent | Pro+ | Coming with 5.3 |
| Eldric Cortex | Reasons over your data. Long-context generation. Domain-specialised experts. | Dual workstation cards OR datacentre OR Eldric-hosted cloud | Enterprise | First specialist with 5.4 |
The same matrix-memory primitive powers all three. Nexus is the recognition layer, Pro is the same with richer per-theme representations, Cortex is the reasoning layer above. You can run all three on the same install; the platform routes work to the tier that's licensed and available.
The recognition layer of the platform. Every conversation, document, and event gets a theme tag without calling a language model — a single matrix multiplication does it in microseconds. Distilled from a frontier model, packed into a small artifact, runs on the CPU you already have. Bundled with every Eldric install.
A worked example. A user message reads:
"My pod is CrashLoopBackOff. Logs show the app crashes with exit code 1 immediately. Adding `command: sleep 3600` and exec'ing in to debug — the app runs fine that way."
Nexus recognises this as devops/container/k8s with high confidence, and tags the conversation in the matrix memory of the customer's installation. Done in roughly the time it takes to read this sentence's first word.
Where it fits: automotive, edge, IoT, mobile — anywhere a GPU isn't available and recognition still has to happen continuously. The Raspberry Pi 5 in a workshop, the SoC in a parked car overnight, the NUC in a clinic's IT closet.
Nexus Pro keeps the matrix-memory primitive but stores multiple representative exemplars per theme instead of one averaged vector. A theme like "Python ML data wrangling" doesn't get collapsed with "Python web backend" because the exemplars are stored separately and matched individually. Result: roughly 100× more themes, sharper disambiguation between close concepts, and a clean novelty signal when a query doesn't match any stored exemplar.
Who it's for: customers with deep, specialised vocabularies. A law firm's case taxonomy. A hospital's clinical terminology. A research lab's project structure. A bank's transaction-class hierarchy. Anywhere "168 generic themes" stops being granular enough.
What it needs: a workstation tier of hardware. A single RTX 6000-class card holds Nexus Pro in memory with comfortable headroom; a dual-card workstation lets it run alongside a small LLM for Layer 3 reasoning without thrashing.
How customers turn it on: a Pro+ tier admin opens the dashboard, selects "Nexus Pro" as the recognition tier, and the platform pulls the appropriate model artifact on activation. The default Nexus stays available as a fallback for any worker that doesn't have the hardware.
Cortex is no longer just a classifier — it's the reasoning layer that generates insight from what Nexus and Nexus Pro recognise. When you need to explain WHY a conversation is novel, summarise a year of context in one pass, or compose a response that draws on the customer's entire institutional history, Cortex is the model that does it.
Built on the same primitive, scaled up. Cortex uses xLSTM-architecture matrix-memory blocks — the same associative recall that powers Nexus and Nexus Pro, stacked into a frontier-scale generative model. Linear attention (not quadratic) means it can read a year of session history in a single forward pass without the context-window blowup that limits traditional transformer-based reasoning models.
Domain-specialised, not generic. Cortex ships as a family of domain experts — automotive, code, medical, finance, legal, science, and so on — combined through a Mixture-of-Experts router. Most queries only activate one or two experts, so effective compute is far less than the total model size. Customers running just one domain (a hospital, a fleet operator, a bank) can license the matching specialist alone and run it on a workstation with dual high-end cards. Customers wanting the full multi-domain reasoner license the assembled model and run it on a datacentre tier or consume it from the Eldric-hosted cloud.
What it's for:
Cortex's domain experts roll out one at a time, starting with 5.4. Each expert ships as a standalone product for customers who only need one domain, BEFORE the full multi-domain assembly is complete. A hospital can buy Cortex Medical without waiting for Cortex Finance. A bank can buy Cortex Finance without waiting for Cortex Legal. Continuous progress, not a single release where everything finally works.
Eldric's dreaming engine runs in three layers — each tractable on different hardware. The tiers map onto the layers:
The first two layers ship on every Eldric install. The third is optional, opt-in, and matches the compute you've decided to allocate. See what's next for the 5.2 architecture preview.
Nexus, Nexus Pro, and Cortex are all built on the same primitive: associative matrix memory with associative updates and single-step recall. Nexus uses it directly with one averaged vector per theme. Nexus Pro extends it with multiple exemplar vectors per theme and compressed softmax retrieval. Cortex uses it as the building block of every xLSTM expert in a Mixture-of-Experts reasoning model. The family scales by adding more memory and more compute — never by switching architectures.
That coherence matters operationally: a customer upgrading from Nexus to Nexus Pro keeps the same theme taxonomy. A customer adding Cortex on top of Pro keeps the same domain structure. The whole family understands the same world; the tiers just understand it at different depths.
Each tier ships in a general variant by default, with domain-specialised variants available for verticals where the vocabulary matters:
Additional verticals — legal, science, industrial, devops, IT, security — land in subsequent releases as Cortex experts. Admins pick variants in the dashboard; switching is a file swap, no service restart.
Modern brains recognise far more than they generate. When you walk into a familiar room you don't compose a description of it — you just know what it is. Recognition is a pattern match against memory. Generation is a different process, and it costs more.
Most AI architectures collapse these two — every classification becomes an LLM call. That works on a workstation with a GPU. It breaks on a Raspberry Pi, on a car parked overnight, on a remote-office server. Eldric's recognition family is the bet that recognition can live separately from generation — and that splitting them is what lets the platform run everywhere.
The bet seems to work. Nexus answers in microseconds where a generation-based classifier would take seconds. The customer who runs a 4-engineer service desk on a NUC gets the same memory and recognition behaviour as the customer who runs a 200-person engineering org on a datacentre cluster — different hardware, same shape of recall. And when reasoning is what the situation actually needs — when a customer wants Cortex to write up why a conversation is unusual, not just that it is — the family scales to that without abandoning what made the recognition layer work.
Customers running just one Cortex specialist can deploy on a workstation tier. Customers wanting the full multi-domain Cortex consume it via Eldric-hosted cloud or self-host on a datacentre tier when the full assembly lands.
See what's next for the broader architecture preview, or documentation for the integration story.