Models & Infrastructure

Intelligence you can upgrade. Hardware you own.

ZenithAI separates the experience your teams use from the models that power it. Four stable tiers run on open-weight models hosted on your GPUs, and can be exchanged as better models are released — without disrupting a single user.

Sec. 01 The tiers

The right engine for every task.

Routing between tiers is automatic — a quick lookup never occupies your deepest model, and a complex analysis is never short-changed by a light one. Users can also choose a tier explicitly per conversation.

I

Instant

Immediate responses for lookups, utilities and speech transcription.

Example class · compact models, 2–5B
II

Fast

The everyday workhorse: general questions, summaries and routine drafting.

Example class · efficient models, 4–9B
III

Smart

Analysis, document reasoning with citations, coding, formal drafting.

Example class · capable models, 12–30B
IV

Genius

The deepest reasoning tier for complex, multi-step analytical work.

Example class · large open-weight models, sized to your infrastructure

Models are examples, not commitments. Each deployment selects models from the current generation of leading open-weight families — such as Google’s Gemma line and comparable open models — matched to your hardware, languages and workload. When the state of the art moves, your tiers move with it.

Sec. 02 Model flexibility

Bring the model that fits. Change it when it doesn’t.

ZenithAI is built on open, standard model-serving technology rather than a proprietary inference stack. That single decision buys your organisation lasting freedom.

  • 01No model lock-in — supported open-weight models can be evaluated and adopted per tier
  • 02Bring your own model — organisation-specific or domain-tuned open models hosted in your deployment, subject to compatibility validation
  • 03Residency without thrash — models are pinned to GPUs with health monitoring, so switching tiers doesn’t mean waiting for reloads
  • 04Local embeddings too — document retrieval uses embedding models hosted on your own hardware, not an external embedding API

Sec. 03 Hardware

Sized to your organisation — not the other way around.

ZenithAI runs efficiently on commercially available GPU hardware. These are illustrative starting points; every deployment is sized during scoping against your user count, languages and workload.

Minimum · x86

Two-GPU server

Two NVIDIA RTX PRO 5000 Blackwell cards (48 GB each, 96 GB in total), 128 GB of RAM and a 1 TB SSD run all four tiers, Instant, Fast, Smart and Genius, with document intelligence and OCR. This is the smallest server supported; buy above it for many concurrent users. Hardware requirements.

Option · ARM64

ARM64 server

An ARM64 server with Ubuntu 24.04 LTS, 16 or more cores and 96 GB of NVIDIA graphics memory (one 96 GB card or two 48 GB cards) runs the same four tiers. Extra cards add headroom for concurrent departments.

Example · Enterprise

Multi-node cluster

Multiple GPU nodes separate interactive tiers from heavy analytical workloads and OCR pipelines, supporting enterprise-wide rollout with capacity isolation.

Efficiency is an engineering discipline here. Tiering, GPU residency pinning and CPU-hosted embeddings extract full value from every card you buy — the reference deployment serves a working organisation’s daily load, including document intelligence and OCR, from a two-GPU footprint.

Questions

Models and hardware, answered.

Which models does ZenithAI run?

Open-weight models hosted on your own GPUs, arranged in four tiers: Instant, Fast, Smart and Genius. The models named on the page are examples, not commitments. Each deployment picks models that fit its hardware, languages and workload.

Can we bring our own model?

Yes. Organisation-specific or domain-tuned open models can be hosted in your deployment, subject to a compatibility check.

How does ZenithAI choose which tier answers?

It routes automatically, so a quick lookup does not occupy your deepest model and a complex analysis is not handled by a light one. Users can also pick a tier for a conversation.

What hardware do we need?

The smallest supported x86 server has two NVIDIA RTX PRO 5000 Blackwell cards (48 GB each, 96 GB in total), 128 GB of RAM and a 1 TB SSD, and runs all four tiers. An ARM64 server with Ubuntu 24.04 LTS, 16 or more cores and 96 GB of NVIDIA graphics memory runs the same tiers. Larger deployments are sized during scoping.

Does retrieval use an external embedding service?

No. Document retrieval uses embedding models hosted on your own hardware.

The next step

Get a sizing conversation, not a sales pitch.

Tell us your user count, languages and isolation requirements — we’ll propose tiers, models and hardware with a transparent cost picture.