Every open-source AI model. One local runtime. Your data never leaves your device.

Mission & vision

Sovereign AI, by default

• Local-first

  Runs entirely on hardware you own. No cloud round-trips, no third-party inference.

• Sovereign by design

  Data residency is absolute: models come to your data, never the reverse.

• Certified & provenance-backed

  Marketplace models will be adversarially tested before listing, and every answer traces back to the expert that produced it.

• Adaptable

  Planned: an AI-assisted pipeline that continuously generates new SLMs, skills, tools, and MCPs.

Prototype demo

Watch the prototype run

Orchestrator v0, running entirely on an 8GB Apple M1 Mac: a question routed across a panel of specialist experts and synthesized into one answer, with per-expert provenance, relevance scores, and real, unedited timings — no cloud in the loop.

Short public preview shown here; a full walkthrough is available to investors and design partners on request.

How it works

A panel of experts, on your machine

Instead of one giant generalist model in someone else's datacenter, NIKITRIA runs a panel of small specialist models on your own hardware and combines their answers — with every step visible and attributable.

1. 01

   Ask

   Your question is processed entirely on your device. Nothing is sent anywhere.

2. 02

   Route

   An embedding-based router matches the question to the best-suited specialist expert — semantic matching, not keywords, and still no network calls.

3. 03

   Panel of experts

   One or more specialist Small Language Models produce candidate answers within a strict on-device memory budget.

4. 04

   Synthesize with provenance

   The answers are combined into a single response, and every part of it traces back to the expert that produced it.

Architecture shown at the conceptual level.

Benchmarks

Honest numbers, or no numbers

We publish only measured values, with full methodology — the way MLPerf-style benchmarks are reported: disclosed hardware, software versions, workload, and a method anyone can re-run. Until our internal benchmark rulings are finalized, every figure below is an explicit placeholder. Nothing on this page is an estimate.

A separate Track B pilot explores 1-bit models (BitNet via bitnet.cpp). Any Track B result will be published separately, with its own hardware context — never blended with the figures above.

Why this is the future

One runtime for all open-source SLMs

The open-model ecosystem is exploding, regulation is pushing AI on-premises, and the economics of small models keep improving. NIKITRIA's bet is that the missing layer is a trusted local runtime that unifies all of it. The evidence below is external research, labeled with its source — separated from anything NIKITRIA claims about itself.

• 2M+

  Open models on Hugging Face

  Hugging Face hosts more than two million open models — the second million arrived roughly 335 days after the first. Qwen2.5-1.5B-Instruct, the base model in our prototype, is among the most-downloaded text models. Open families now include Qwen, Phi, Gemma, Llama, SmolLM, and BitNet.

  Hugging Face, 2025

• 10–30×

  SLMs are the future of agentic AI

  An NVIDIA Research position paper argues small language models are sufficiently powerful, inherently more suitable, and necessarily more economical for many agentic tasks — and that heterogeneous, multi-model systems are the natural architecture. It reports that serving a ~7B SLM is 10–30× cheaper in latency, energy, and FLOPs than a 70–175B LLM. That is the panel-of-experts thesis.

  Belcak et al., NVIDIA Research, arXiv:2506.02153, 2025

• On-device

  The platform shift is already underway

  Apple Intelligence (on-device models plus Private Cloud Compute), Microsoft's Copilot+ PCs and Phi family, and Google's Gemma and Gemini Nano all point the same direction: inference is moving onto the device.

  Public product announcements, 2024–2025

• $15–$30

  Per million output tokens, frontier cloud

  Frontier flagship models charge on the order of $15–$30 per million output tokens. A capable free on-device floor pushes routine inference toward zero marginal cost.

  Published provider price lists, 2025

• €35M / 7%

  Regulation makes sovereignty a requirement

  India's DPDP Act 2023, the EU AI Act, and GDPR are turning data locality from a preference into an obligation. The EU AI Act's Article 99(3) sets fines for prohibited practices of up to €35M or 7% of total worldwide annual turnover, whichever is higher — above GDPR's €20M / 4% tier.

  EU AI Act Art. 99(3); DPDP Act 2023; GDPR

• 6× / −82%

  The 1-bit efficiency frontier

  Microsoft's BitNet b1.58 and bitnet.cpp run 1-bit LLMs on CPUs, with Microsoft reporting up to 6× faster inference and 82% less energy than conventional approaches — reinforcing the modest-hardware thesis and the direction of our exploratory Track B pilot.

  Microsoft, BitNet b1.58 / bitnet.cpp

All figures in this section are external research, quoted with source and date. None are NIKITRIA measurements.

Business model

Five pillars, one adaptable platform

The model we are building toward. Every pillar below is a plan under active development, not a shipped product — the working prototype today is the orchestration runtime they all sit on.

1. 01

   Certified SLM marketplace

   A marketplace of specialist models, each adversarially tested before certification, with creator-friendly revenue sharing to build the supply side.

2. 02

   MCP, skills, tools & apps marketplace

   The same certification and distribution rails extended to MCPs, skills, tools, and SLM-powered apps.

3. 03

   Hosted inference tier

   Usage-metered inference in a customer's VPC or on managed servers, for teams that want the platform without operating it.

4. 04

   BYO-server enterprise tier

   Software licensing on the customer's own infrastructure — the sovereign-AI wedge. Data never leaves their perimeter.

5. 05

   Integration services

   Hands-on integration as the cold-start go-to-market, productized over time into SDKs.

BYO-server · on-premise

Your servers. Your models. Your data.

The planned enterprise tier licenses the entire NIKITRIA platform to run on infrastructure you already control — your servers, your VPC, your air-gapped rack. No cloud dependency, no vendor lock-in, and data that never crosses your perimeter.

• Data sovereignty, literally

  Models, routing, and synthesis all execute inside your network boundary. There is no telemetry path to us.

• No lock-in

  Built on open-source model formats (GGUF) and open protocols — your models remain yours, portable by construction.

• Regulatory alignment

  Designed for DPDP, GDPR, and EU AI Act environments where off-premises inference is a compliance risk, not a convenience.

In-company fine-tuning

Your data, distilled into your own private experts

A planned NIKITRIA service: fine-tune a company's or industry's proprietary data into domain-specific Small Language Models that run on the customer's own hardware — private by construction, because the data and the model never leave the building.

• Specialists beat generalists at the job that matters

  A model trained on your processes, vocabulary, and edge cases — retaining the context generic pilots lose.

• Fast, cheap iteration

  NVIDIA's SLM research notes small models can be fine-tuned overnight rather than over weeks, and are 10–30× cheaper to serve than frontier-scale LLMs.

• Private by construction

  Training data, adapters, and inference all stay on infrastructure you control.

Offered initially as an integration-services engagement (pillar 5), productizing over time.

Education vertical

Curriculum-specific AI for India's classrooms — offline and affordable

Indian institutions are being asked to teach AI at national scale on constrained budgets. NIKITRIA's planned education offering pairs curriculum-specific SLMs with hardware schools already own: no GPU cluster, no per-student cloud bill, and it works with no internet at all.

• Curriculum-specific experts

  Planned: specialist models aligned to specific syllabi and courses, for student practice and faculty support.

• Runs on existing computer labs

  The same local-first runtime shown in our prototype — designed for modest, consumer-grade hardware rather than GPU clusters.

• Offline by design

  Connectivity gaps stop being a blocker: everything runs on the machine in the room.

Market opportunity

A market measured in ranges, not cherry-picked numbers

Market-size estimates for small language models diverge 3–5× across research firms, so we present the range with each firm named and dated — and we separate external projections from anything NIKITRIA claims about itself.

All market figures are external vendor research — projections, not audited facts — cited with firm and date.

Competitive differentiation

Everyone wraps the same engine. Nobody integrates the stack.

Ollama, LM Studio, Jan, and GPT4All all wrap the same llama.cpp engine — raw token throughput between them is close to a wash. They differ on packaging, GUI, and API surface, not fundamental capability. None offers a certified marketplace, panel-of-experts orchestration, full provenance, and a BYO-server sovereign tier as one integrated product. That combination is what NIKITRIA is building: orchestration and provenance run in the v0 prototype today; the certified marketplace, certification, and BYO-server tier are planned.

Roadmap

Where this goes — labeled honestly

One thing on this roadmap exists today; everything else is a plan with a target, and is labeled that way.

1. Working today

   Orchestrator v0 prototype

   Panel-of-experts routing, synthesis with provenance, fully local — running on an 8GB Apple M1 Mac, self-capped to 4GB of RAM.

2. Target: 10–12 months

   Certified-marketplace beta

   Certified specialist SLMs with adversarial testing and creator revenue share — the core milestone of the pre-seed plan.

3. Planned

   Hosted inference tier

   Usage-metered VPC/server deployments for teams that want the platform managed.

4. Planned

   Enterprise BYO-server GA

   General availability of the sovereign licensing tier on customer infrastructure.

5. Exploratory

   Track B: 1-bit models

   A pilot exploring BitNet 1-bit models via bitnet.cpp for even lower-footprint inference. Results, if any, will be reported separately with their own hardware context.

Founder

Who is behind NIKITRIA