SwypikBuilt in EuropeInvestors
Ilaria · Native AI

Not one model in a datacenter. A living network.

Ilaria is the native AI model and distributed cognitive system for SwypikOS. It is one model family, IMC, designed to be trained from scratch, and it is designed to live on every device as a cell that learns from verified experience and learns which peers are worth asking.

Explore Myriad See what’s built
IMC · no pretrained weightsNative ternary 1.58-bitSub-byte accelerationMemory local by default
1 Billionparameters in IMC-1B, deployment target
1.58 bitsnative ternary weights: −1, 0, +1
~10×smaller weights vs FP16 (theoretical)
Localzero cloud latency, total privacy
01 · The model

One genome. Four sizes. Zero borrowed weights.

IMC (Ilaria MicroCortex) is the only model family in Ilaria. No external pretrained language model is part of the architecture. The smaller sizes are scientific checkpoints of the same design, not separate products: each rung validates the architecture before the 1B run.

Scale ladderexact parameter counts
  1. IMC-125MFirst scale gate125M parameters
    Validated on ternary and full-precision paths
  2. IMC-250MValidation250M parameters
    Same architecture, next rung×2.00
  3. IMC-500MValidation500M parameters
    Unlocked after network gates pass×2.00
  4. IMC-1BDeployment target1 Billion parameters
    One full resident model per eligible device×2.00
IMC-1B specificationcanonical architecture
Architecture
Proprietary decoder-only transformer
Target Scale
1 Billion parameters (IMC-1B)
Precision
Native ternary 1.58-bit {-1, 0, +1}
Activations
Low-precision int8
Context
2,048 tokens (extensible)
Attention
Grouped Query Attention (GQA)
Normalization
RMSNorm · Rotary Positional Embeddings
Embeddings
Tied embedding architecture
Weights
Ternary block projections
Tokenizer
Proprietary IlariaLex tokenizer
Runtime
On-device sub-byte execution engine
Intellectual Property
100% Proprietary · Closed-source
02 · Ternary native

Sub-byte weights: −1, 0 and +1.

IMC’s block projections are ternary during training itself, not squeezed down afterwards. With native 1.58 bits of information per weight and int8 activations, our proprietary on-device packing targets roughly 10× smaller weights than 16-bit models.

Input20 ternary weights
Proprietary sub-byte packing demo
  1. −10
  2. 01
  3. +12
  4. −13
  5. 04
  6. +15
  7. −16
  8. 07
  9. +18
  10. −19
  11. 010
  12. +111
  13. −112
  14. 013
  15. +114
  16. −115
  17. 016
  18. +117
  19. −118
  20. 019
Sub-byte ternary encoding · 1.58 bits / weightPacking weights into 32-bit registerregister value1,475,178,015
OutputCompact 32-bit register
0x57ED6E1FPacked bytes 1f 6e ed 57
register value High packing density (>99%)32-bit hardware word
Verified · Sub-byte packed

~10× Smaller Weights

Theoretical ~10× smaller weights (1.58-bit vs 16-bit), the basis for the target of fitting 1B-class models inside consumer device RAM.

Signed Checkpoints

Signed checkpoints, verified on load: modified or unsigned weights are rejected before they run.

Zero Floating-Point MACs

Replaces power-hungry matrix multiplication with efficient integer additions, dramatically extending device battery life and reducing thermal throttling.

Proprietary on-device compression engine. Optimized for mobile and desktop hardware acceleration.

03 · Myriad

A network that learns its own shape.

“At one million installations, Myriad is one million situated cognitive cells, not one million dumb replicas.” Seven research pillars make that possible, each one checkable on its own.

Pillar 01Contracts built

Genome / Phenotype

Every cell shares a stable Genome. Each develops its own Phenotype from what it experiences.

  • Genome: IMC architecture version, IlariaLex tokenizer, protocol version, Genesis ancestry, model-format contract
  • Phenotype: specialized weights and checkpoint lineage, local memory, competence vector, device embodiment, learned peers, trust
One architecture becomes many useful experts without incompatible model families.

Built means implemented and verified at tiny, mechanism scale. Behavior at network scale (multi-expert network gain, real device transfer, million-node swarms) has not been demonstrated yet.

04 · Hippocampus

Learns a fact once. Keeps it local.

In the verified memory workflow, Ilaria learns a fact from a single exposure. Personal hippocampal memory stays on the device by default, outside the reach of global training.

This device
HippocampusEpisodicStore
  • Thursday pickup · 16:30
no implicit access
Global trainingMyriad · Compute Fabric workers
  • Training workers are spec-forbidden from personal.hippocampus, user files and secrets.
  • Experience that does travel goes as signed capsules, not raw private data.
  • Remote inference in the Swypik app pilot needs its own explicit consent.
IlariaHippocampus · on this devicelocal

  1. Remember: Thursday school pickup is at 16:30.

  2. Got it. I’ll keep that.

  3. memory.write 1 episode · stored on this device
  4. two weeks later

  5. Can I take a call Thursday at 16:15?

  6. memory.read 1 episode recalled

  7. That would run into your 16:30 school pickup. A slot at 15:30 keeps you clear.

Illustrative flow of the memory workflow, not a transcript of model output.

Cryptographic Memory BoundaryHardware-enforced episodic privacy
  • Zero Ambient Access: Background workers and training jobs cannot read personal episodes.
  • Local-First Storage: User interactions and memory graphs remain strictly on the host device.
  • Explicit Capability Tokens: Memory write/read requires short-lived user-consented grants.
05 · Evidence

Experience transferred. Nothing forgotten.

Proof-Carrying Experience Transfer v2 asks one question: can a cell learn new skills from another cell’s signed experience during Collective Sleep without losing what it already knew? A matched control receives the same budget with a deliberately wrong mapping.

Formal promotion gate · PASSControlled synthetic benchmark
+27.08 ppmean top-1 causal advantage over the matched control
4 / 4seeds with positive NLL advantage (mean +1.4159)
0 ppknown-anchor drop on every seed
0seeds with a top-1 regression (3 / 4 positive)

Per-seed top-1 advantage, not averaged away

+83.33seed 7
+8.33seed 11
0.00seed 19
+16.67seed 23

12 synthetic new skills, 8 anchor skills, a tiny IMC, frozen SHA-256 dataset identity, gate thresholds unchanged. This shows the mechanism works; it says nothing yet about model quality. Next step: repeat on real IMC-125M with 5–10 seeds.

How a cell teaches a cell

  1. Verified experienceCell A acts and a verifier confirms the result
  2. Signed capsuleEd25519 signature, ancestry and privacy class checked before replay
  3. Collective SleepCell B replays the capsule alongside known anchors with a frozen-teacher term against forgetting
  4. Promotion gatePromote the winner or promote none
Awakeinference · world · memory · experience
Sleepreplay · consolidate · test forgetting · exchange
06 · Training & data

A data chain that fails closed.

No production token is trained until every source is pinned, licensed, hashed and reviewed by a person. The pipeline is built end to end and rejects anything that hasn’t cleared review. Automatic approval: never.

> 3,000,000curated candidate documents
> 5 GBclean candidate text
~1.3Bconservatively estimated tokens
1Btoken Genesis horizon for IMC-125M
Genesis curriculumBalanced domain distribution
  • General knowledge30%
  • Software & Code22%
  • Mathematics & Logic15%
  • Science & Technical11%
  • Systems & Hardware10%
  • Agent & Tool trajectories8%
  • World & Embodied dynamics4%
Training probeIMC-125M · ctx 2048
  • InfrastructureEnterprise AI accelerator
  • ExecutionMixed precision + gradient checkpointing
  • Target horizon1B tokens Genesis curriculum
  • PipelineDeterministic checkpointing & resume

Highly optimized training memory footprint, running with full gradient checkpointing and custom optimization steps.

Production chain

  1. Source lock
  2. Rights evidence
  3. Human review
  4. Tokenizer freeze
  5. Curation & encoding
  6. Dataset manifest
  7. Preflight
  8. Pretraining launch
built & passedin progressnext

Supported by

  • EuroHPC Joint Undertaking

    Compute allocation awarded by the European High-Performance Computing Joint Undertaking — MareNostrum 5 ACC (H100)

07 · Honest status

Built. Next. Vision.

The architecture, training pipeline, promotion gates and data chain are built. Full pretraining comes next. Until the evaluation suite has spoken, we make no claims about model or chat quality.

Built

The machine that makes the model

  • Self-contained IMC architecture; 125M-parameter scale validated across ternary and full-precision paths
  • Deterministic checkpoints, exact resume, chunked loss, gradient checkpointing, distributed execution
  • Proprietary on-device packing codec and zero-heap CPU reference engine
  • Signed PCE with cryptographic verification, learned Connectome and top-K router, Collective Sleep v1
  • PCE Transfer v2 promotion gate passed on benchmark suites
  • Rights-gated data chain, IMC-125M preflight and infrastructure probe
Next

Train, then prove the network

  • Close the human rights review, then freeze the production IlariaLex tokenizer
  • IMC-125M pretraining on 1B tokens, with matched ternary and full-precision controls
  • Scale PCE transfer on real IMC-125M across benchmark seeds
  • Prove learned-synapse routing across multi-expert domains
  • Scale to 250M and 500M only after network and continual-learning gates pass
Vision

One IMC-1B on every SwypikOS device

  • IMC-1B Genesis, to be trained from scratch once IMC-125M passes its gates
  • A full resident model on every eligible device, each with its own memory, lineage and competence
  • Millions of cells joined by a learned, verified connectome
“This is a hypothesis to prove, not a marketing assumption.”
08 · Ecosystem

Where Ilaria lives.

SwypikOS

Ilaria is SwypikOS’s native AI. It verifies signed effect receipts but holds no OS authority of its own: effects go through explicit capabilities.

Explore SwypikOS

Swyp

Ilaria’s canonical spec is a Swyp file. A swyp judge tool lets a model propose code and get a verdict back; a run with the live model is next.

Explore Swyp

Swypik app pilot

An “Ask Ilaria” tab in the mobile pilot: foreground only, explicit consent for remote inference, contribution off by default. Frozen for review, not published.

How it all connects
One protocol between cognition, language and authority.Ilaria thinks, Swyp verifies, SwypikOS authorizes.
See the Nexus

Try “open movies”, “open go”, “stiri” or “investors”.