Research

Smaller models.Better frameworks.Verifiable memory.

Why Meotis is building intelligence that runs on the Internet Computer, and why this programme is open to researchers and developers.

With the ecosystem
DFINITYORIGYN

Already measured, already shipped

2.9×more inference from software alone, same weights, same mainnet budget
40 bninstruction units per call: the deterministic ICP budget, not negotiable
48 hbetween our open publication and the v0.12.0 release by OnicAI

Five research tracks

  1. Execution engine

    Fewer instructions per useful operation: vector maths, quantisation, ternary models and session optimisation.

  2. Models

    Small models specialised by task, evaluated under the real constraints of ICP, not only on GPU benchmarks.

  3. Execution frameworks

    Orchestration, memory routing, tool use, verification and multi-call strategies under strict budgets.

  4. Proof

    Every critical agent action gets an explicit budget, an execution receipt and a traceable memory path.

  5. Cognitive heritage

    Encrypted local vaults, certification through ORIGYN SDK and controlled access to authenticated human memory.

The AI industry is no longer only chasing large models. It is chasing more intelligence per unit of compute. Phones demand efficiency for privacy and energy; blockchains demand it for consensus and verifiability. Both point the same way.

Efficiency is the product

On ICP, a call runs under a deterministic budget: you cannot buy compute during execution. We ran the same model, at 500 million parameters, with the same weights and the same mainnet limit, through two different software paths: 10 tokens per call → 29 tokens per call. No new hardware, no larger model. The model is the engine; the execution framework is the vehicle.

An open research loop

We published the paper, the code, the raw measurements and the reproduction steps. Within 48 hours, the maintainer of llama_cpp_canister (OnicAI) traced the bottleneck independently, confirmed it on mainnet and shipped a release ~2.8× faster. Open research became working infrastructure within days.

Measure. Publish. Reproduce. Improve.

Memory as heritage

Agents do not only need a brain: they need a memory they can trust. We call cognitive heritage the evolving body of knowledge, decisions and creations a person builds up over a lifetime, today fragmented, with uncertain provenance. Private content stays local or encrypted; ICP anchors authenticity, provenance and permissions; ORIGYN SDK certifies the rights. Human memory becomes a sovereign, programmable and lasting digital heritage.

Paper and artefactdoi.org/10.5281/zenodo.20607598Discussion and reproduction on mainnetforum.dfinity.orgArtefact repositorygithub.com/SimlowkerOnicAI release · llama_cpp_canister v0.12.0github.com/onicai

Join the research programme.

Small models, agent frameworks, WebAssembly, quantisation, provenance, private memory or decentralised AI: if you build in one of these fields, this programme is yours.

Let’s talk about your company.

  • 30 minutes, by video or in Neuchâtel
  • With the founding team
  • No strings attached

Meotis, The Cognitive Heritage

Write to hello@meotis.chDescribe my project

Meotis

Discovery call

  • 30 min
  • Video, link sent by email

Choose a day

Choose a time

Neuchâtel time

Pick a day, then a time.