METIS turns a pretrained model into your model. Knowledge is organised and retrieved by the Gluon Geometric Knowledge Layer; the model is adapted and continually updated by the F-HKAN Runtime with CSPR Adapter — and every change is measured against thresholds committed before the run.
Unified geometric index across models and modalities: grounding and retrieval at sub-millisecond speed, and a drift gate watching the distribution.
The knowledge layer →Runtime, adapter, consolidation: a pretrained model adapted to your domains, learning continually — with inference speedup, without full retrains.
The model layer →Pre-committed thresholds, before/after evaluation on every adaptation, retention tracking, version lineage, drift monitoring. The other planes stand on this one.
Trust & Method →Adaptation without continual learning and without speedup — and a retrain from scratch on every knowledge update.
Faster inference on a model that never learns your domain. If that is all you need, it is the right tool — and we say so.
Retrieval without adaptation — the model itself stays generic, whatever the knowledge base learns.
We think you should know what a vendor won't do before you trust what it will. The anti-claims page is a differentiator here, not fine print.
Knowledge organised, output generated, new knowledge integrated, everything measured. A METIS deployment is that loop running on your hardware, inside your perimeter — with the evaluation record to prove what it did.
The pretrained backbone stays frozen; adaptation runs beside it, on your GPUs — against acceptance thresholds committed with your team before the run.
The adapted model serves on spline kernels — faster inference on your domains, grounded and accelerated again by the knowledge layer.
New knowledge integrates online — no full retrains. Prior domains are protected during consolidation; the base grows only through the task-success gate.
Before/after evaluation on every update; retention tracked; version lineage recorded; drift monitored. The evaluation record ships with the model.
On-prem and air-gapped deployment is a first-class story, not an afterthought. Your corpus, your model, your audit trail — METIS does not train on your data beyond your perimeter.
A briefing covers your domains, your deployment constraints, and the thresholds we would commit to before any run.
METIS is developed and operated on production-scale multilingual analytical workloads.
What the system is made of, at capability level. Two planes do the work — knowledge and model — and both stand on a third: measurement. How exactly each plane does it is a technical briefing under NDA.
drawn as the foundation because it is one — the competitive axis, made visual
"The knowledge plane organises and indexes. The model plane generates and learns. The planes meet at a measured boundary — not a marketing seam. We measured the boundary ourselves, and we engineer on the right side of it."
Two technologies in synergy. F-HKAN is the runtime — spline-based computation with custom GPU kernels; it delivers speedup and parameter efficiency on its own. CSPR is the continual-learning layer on top — domain experts, slow consolidation, memory protection. Together they turn a pretrained model into yours, and keep it learning.
Adaptation runs as a parallel adapter beside a frozen pretrained backbone — an addition, not surgery on your weights. The expert path learns and consolidates; the fast path serves inference.
The pipeline probes your backbone, configures itself from measured evidence, and never leaves your infrastructure. There is no zero-shot speedup by construction — the adapted model is earned by this run.
New knowledge is integrated online, on a cadence — without full retrains. Prior domains are protected during consolidation, and retention is measured and reported for every update: the system tracks what was kept, not just what was gained.
Continual learning with measured retention — trade-offs are measured and disclosed, not denied.
Continual learning and audit now extend to mixture-of-experts backbones — MoE is supported on the CL and audit axis, with the same step-0 identity, retention and router-health gates. All speedup figures on this page are measured on dense backbones; MoE carries no performance figures, measured or projected.
A before/after evaluation across reasoning, instruction-following, and language quality — against thresholds committed before the run.
Public materials carry capability claims and measured results. Consolidation schedules, protection mechanics, and implementation numerics are covered in a technical briefing under NDA.
The METIS knowledge layer: a unified geometric index built from the internal structure of many independently trained AI models — text, vision, audio, generative — projected into one shared low-dimensional space with a stable cluster topology. New models and modalities dock into the existing structure rather than fragmenting it.
The layer organises and indexes knowledge. Generation is always done by the model. That is a trust statement, not a limitation — the two layers meet at a measured boundary, not a marketing seam.
Query → geometric lookup → relevant concept context injected. Measurably faster, more factual responses than an ungrounded baseline.
New modalities align into the shared space over frozen encoders in minutes-class time — the geometry supplies the prior.
Existing knowledge is recalled reliably — zero spurious duplication in testing. Adding knowledge is gated on task success, never on a similarity score. Nothing enters the base by accident.
Geometric off-manifold detection as a service function: the geometry flags when incoming distribution drifts off the known manifold, and feeds the measurement plane.
The geometry doubles as an instrument: it shows where representation structure is strong and where the underlying data was thin.
The visual asset of this site is the data itself: renderings of the actual cluster and manifold structure of the unified index. Commissioned stills, optionally a slow-rotation render — never a generic particle swarm.
7 model families · ~1.1T parameters · 1,003 clusters · 100% absorption, zero new clusters
vs unmodified baseline over 99 queries · no quality loss — wins forced-choice judging
sub-millisecond retrieval on the shared geometry
in ~30 minutes, over frozen encoders
hold text + vision + audio representations together
matched random-vector control finds no structure where the real index finds 1,003 clusters
Geometry recalls. Task success gates capture. Retrieval is a geometric operation; admission to the knowledge base is earned by measured task outcomes.
"Gluon uses geometry as a map, not an engine: it organises and retrieves knowledge at sub-millisecond speed, while generation stays with the language model."
The Gluon layer does not generate content, does not out-route a learned router, and does not curate meaning. Every METIS function that requires judgment runs in the learned model layer — which is exactly what the F-HKAN and CSPR side provides. We measured the boundary ourselves, and we engineer on the right side of it.
The same composed system — adapted model, geometric grounding, measurement plane — instantiated wherever a knowledge base changes faster than a retrain cycle and answers must be auditable. One pattern runs at production scale today; the other five are applicability, stated as such.
A continuous multilingual open-source stream — high volume, mixed quality, always moving — and analysts who need structured fields and reports, not raw text. Freshness is the product: yesterday's model misses today's entities.
Re-training cycles replaced by measured incremental integration. Grounded responses measured 49.6% faster than an unmodified baseline (99 queries, no quality loss) — and every analytical field traces to a model version with its evaluation record.
METIS is developed and operated on production-scale multilingual analytical workloads.
Precedent and regulation drift continuously; a static model cites last year's law. Drafting support has to ground every claim — and stay inside the firm's perimeter.
Precedent shifts absorbed on a measured cycle instead of an annual model refresh; grounded drafting support inside your perimeter. Qualitative until your corpus is measured — and we say so.
Literature moves weekly; terminology differs by specialty; PHI cannot leave the building. The model has to absorb the new without forgetting the established.
Literature velocity absorbed without from-scratch retraining; the on-prem PHI perimeter is a first-class deployment, not an exception granted reluctantly.
Regimes change; filings and regulatory streams never stop; and model governance asks for artifacts most vendors cannot produce.
Model governance gets what it actually asks for — version lineage, evaluation records, thresholds committed before runs — while the model keeps pace with the regime.
Disconnected networks and hard accountability: which model version produced this output, and what did it score before deployment? Multilingual OSINT is often the adjacent workload.
Air-gapped is the design point, not a degraded mode. Every fielded version is accountable to its evaluation record.
Institutional knowledge decays silently: the base grows stale, the model answering on it stays generic, and nobody can say what the system actually knows.
The base stops decaying and the model keeps up with it — one loop, without a retrain budget line every quarter.
A briefing walks the loop on your corpus, your cadence, your perimeter — and ends with the thresholds we would commit to before any run.
The five arcs that once pointed forward now run in production. Every node on this map is a shipped capability, not a promise — delivered, operating across deployments, and extended to the frontier. Click a node to read what it does.
Knowledge-layer drift semantics feed fleet operations; retention science feeds the production runtime. Where you fit into the plan is a briefing conversation.
This page is written once, carefully, and changes rarely. It is how we work — and it is the reason a compliance officer, a CTO, and an investor can read the same site and trust the same numbers.
Every architecture change is gated on a cheap simulation against a learned baseline before it earns GPU time.
Pass and kill criteria are written down before each run — not fitted to the results afterwards.
Results that close a claim are reported with the same prominence as wins — in the same papers.
Geometry as a generative engine — closed. Geometric novelty curation — closed. Geometry replacing a learned router — no gain. That is why the Gluon layer is sold as a map, never as an engine.
We think you should know what a vendor won't do before you trust what it will.
Adapting a model to your domains is a real training investment. That investment is made once, then amortised: the adapted model runs faster and keeps learning without full retrains. If all you need is a faster static model, modern quantization and serving stacks are the right tool — and we'll tell you so.
Adaptation targets your domains. Every adaptation ships with a before/after evaluation across reasoning, instruction-following, and language quality, measured against thresholds we commit to before the run — and we never move a threshold after seeing a result.
The Gluon layer organises, indexes, and retrieves — reliably and fast. It does not generate content and it does not judge meaning; the adapted model does that. We measured the boundary ourselves and we engineer on the right side of it.
Every measured figure on this site traces to a specific run with a fixed configuration. Where we show a projection, it is marked as one, it is anchored to measurements we actually made, and we'll show you the anchor. Negative results in our research are reported with the same prominence as positive ones — ask us for them.
METIS deploys on your infrastructure. Your corpus, your model, your audit trail.
A projection starts from at least one measured anchor run — interpolation between anchors where possible, near-range extrapolation otherwise. It is marked projected at every occurrence, in a visibly different style from measured numbers. Each one carries an internal provenance note — the anchor runs and the extrapolation basis — that survives due diligence. We never project onto architecture families we haven't measured; subjects without a base measurement carry no numbers at all.
dense transformer backbones · ~1B–14B class · 8 FFN families
Stated as supported scope, not as apology. All published speedup figures are measured on dense backbones; configurations available on request.
Mixture-of-experts backbones are supported on the continual-learning and audit axis — carrying no performance figures, measured or projected.
Want the runs behind the numbers — including the negative ones?
METIS is built by the technical R&D division of creativeplace.io and qs-lab.io — Israel. Officially NVIDIA-affiliated as an AI systems developer. The evidence discipline on this site is how we worked before it was a website.
One system, two technology layers: the F-HKAN Runtime with CSPR Adapter and the Gluon Geometric Knowledge Layer — on a measurement plane.
Thresholds before runs. Negative results reported. Measured numbers, marked projections — nothing unmarked. The Trust & Method page is the company, written down.
METIS is developed and operated on production-scale multilingual analytical workloads.
Formal analysis and experimental validation of the dual-path runtime, with a full claims scorecard — passes and failures both.
An evidence-only account of weight absorption, geometric retrieval and knowledge lifecycle — including the claims our own experiments closed, stated with the same prominence as the wins.
Briefings are led by the team that builds the system — not a sales layer.
Enterprise engagements start with a conversation: your domains, your deployment constraints, and the thresholds we would commit to before any run. Technical depth beyond the capability level is covered under NDA.
METIS is built by the technical R&D division of creativeplace.io and qs-lab.io — Israel. Officially NVIDIA-affiliated as an AI systems developer.
Your email client should have opened with the request addressed to {{ email }}. If it didn't, write to us directly at that address — it goes straight to the R&D team.
