CAPABILITY 01
A trillion parameters that know how to stay quiet.
Roughly 18 to 28 billion parameters activate per forward pass, drawn from a 1T+ pool. Larger capacity, smaller footprint, longer thought.
AbteeX AI Labs — Aotearoa New Zealand
Sovereign
intelligence,
held in the light.
LumynaX (Te Mārama) is the sovereign intelligence platform developed by AbteeX AI Labs in Aotearoa New Zealand. A trillion-plus parameter mixture-of-experts that dynamically activates only what the task requires.
Built to carry scale without carrying waste. Built in a place with a long memory. Built to answer to the people it serves.
Now in private access
Architecture
MoEmixture of experts
Parameters
1.05Ttotal
Active / pass
18–28Bdynamic
Context
512ktokens
§ Platform · 01
Scale, activated only where it matters.
Most trillion-parameter systems light the whole room to read a single page. LumynaX doesn't. The mixture-of-experts architecture routes each forward pass through a small, task-appropriate committee — the rest of the model stays dark and accounted-for.
CAPABILITY 02
Long context, carried steadily.
A 512k window that does not forget the opening line by the time it reaches the closing argument.
CAPABILITY 03
Sovereign by construction.
Data residency, governance, and model provenance are not a checkbox. They are how the system is shaped from the first layer.
CAPABILITY 04
Alignment held in place.
Refusal behavior, tool permissions, and audit trails are continuous properties — not prompt-layer suggestions.
CAPABILITY 05
Built for production, not demos.
Determinism envelopes, streaming budgets, structured outputs, and evaluation harnesses ship with the platform — because the interesting work begins after the demo is over.
§ Architecture · 02
Four layers. One forward pass.
From prompt to provenance in under a second — and every step on the record.
Routing, selection, alignment, execution.
A request descends through four layers. At any moment only a small fraction of the model is lit; the rest stays dark and accounted-for. Nothing is invoked without a reason, and nothing is hidden from the record.
LATENCY p50
340 ms
340 ms
THROUGHPUT
12.4k tok/s
12.4k tok/s
EXPERTS / POOL
128 of 2,048
128 of 2,048
AUDIT
always-on
always-on
Router · attentionL01
Expert selection · 18–28BL02
Alignment envelopeL03
Execution · provenanceL04
ACTIVATION · live experts · 32 of 2,048 shown
§ On sovereignty
Intelligence
made
somewhere
specific,
by
someone
accountable,
answers
differently.
Ko te mārama te tūāpapa — the light is the foundation.
§ Principles · 03
The commitments behind the weights.
A model line carries the decisions of the people who shaped it. These are ours — written down, on the record, and open to challenge.
01 — Provenance
A return address on every weight.
Every parameter descends from a documented predecessor. No orphan checkpoints, no silent retrains.
02 — Consent
A corpus we can name out loud.
Training data is catalogued, licensed, and consented — including the sources we chose to decline.
03 — Residency
Compute with a postal code.
Customer data and inference can stay inside a specific jurisdiction by default, not as an add-on.
04 — Alignment
Safety that answers to a place.
Refusal and policy are shaped to a specific legal and cultural context — not to a generic global average.
§ Operational layer · 04
Meet the layer that works through LumynaX.
LumynaX is the intelligence platform. WeftlineAI is the agentic operations layer teams actually work through — governed execution across channels, business systems, and approvals, all running on LumynaX underneath.
Visit WeftlineAI →Product on LumynaX
WeftlineAI
Agentic operations layer
Shared team memory, approval-aware actions, and controlled execution — deployed to cloud, private cloud, or on-prem. Built for real operating teams, not one-person chat sessions.
Slack · Teams
Governed
Traceable
weftlineai.com →
Stand in the light.
Deployment plans, research partnerships, and sovereign-stack briefings — directly with the team that builds LumynaX.