Home SOWE Network Infrastructure Enigma LYRA Music VST Shoppe
All systems online

Starship No. 27·Sovereign Compute

N27 · Proprietary Sovereign MoE

Enigma

Our flagship sparse Mixture-of-Experts — trained end-to-end on infrastructure we own, served on nodes we operate. Closed weights, sovereign path, no third-party inference. Pre-trained to ~1.5–2T tokens, then aligned with SFT and DPO for multi-agent coding, deep step-by-step reasoning, and production workloads that cannot leave your stack. Parameter counts and weight footprints stay unpublished.

Sparse MoE
Flagship · Own Stack
Top-2
Enigma-Gate Routing
Proprietary
Closed · Owned Weights
Sovereign
Served on Own Nodes
LIVE OPS Enigma in production · Own nodes · Closed weights

Enigma-Gate Router

Hover an expert. Telemetry shows what triggers its dispatch and how the gate routes it.

Gating Telemetry · Sparse MoE
Active Expert
Top-2 Token Router (Enigma-Gate)
Dispatch Slot
Shared Attention + Expert FFNs · Top-2 Dispatch
Trigger Tokens
Awaiting query tokens…
Weights
Trained from scratch · Closed weights
N27-Enigma-Gate initialized.
Trained-from-scratch weights loaded · sovereign path online.
Hover an expert to inspect gating flow…

Specialized Lobes · One Gate

Enigma-Gate routes every token to its top two specialists — domain lobes for code, math/science, general web, and reasoning, ringed by the cognitive operators. Hover any card or node to sync the map and telemetry.

Domain Lobes
Expert 1 · code
Code Expert

Algorithms, systems code, refactoring, compilers, debugging.

Expert 2 · math & science
Math & Science Expert

Proofs, derivations, physics, olympiad problems, quantitative reasoning.

Expert 3 · general web
General Web Expert

Broad factual knowledge, multilingual understanding, general web corpus.

Expert 4 · reasoning
Reasoning & Instruction Expert

Deep chain-of-thought, planning, tool use, instruction following.

Ring 1 · Core Infrastructure
01 · physicist
Physics Core

Physics, applied math, thermal stress, spectral analysis.

02 · echoflux
EchoFlux Coordinator

Mesh sync, edge routing, telemetry streams, heartbeats.

03 · rag
RAG Catalog

Indexed retrieval, corpus grounding, structural validation.

04 · hephaestus
Hephaestus Forge

Zero-destructive fabrication, Rust builds, compiler paths.

05 · creator
Creator Core

Interface design, synthetic layouts, aesthetic blueprints.

06 · operator
Operator Core

Process execution, daemons, cluster reloads, containers.

07 · mage
Mage Core

High-dimensional coords, field equations, warp metrics.

08 · surgeon
Surgeon Core

Telemetry → clean Node, Rust, or Bash execution.

Ring 2 · Cognitive Augmentation
09 · cognitive
Cognitive Augmentation

Relational schemas, decision paths, multi-step logic.

10 · resonance
Emotional Resonance

Tone, sentiment, conversational tempo adaptation.

11 · context
Contextual Awareness

Long-horizon facts, dialogue threads, memory buffers.

12 · synaptic
Synaptic Amplification

Feedback tuning, scale factors, reasoning thresholds.

Ring 3 · Operational Disciplines
13 · commander
Strategic Commander

Tactical planning, risk profiles, high-stakes ops.

14 · medic
Bio-Medical

Cellular biology, genetics, molecular synthesis.

15 · professor
Academic Professor

Peer-reviewed synthesis, long-form academic prose.

16 · engineer
Physical Engineer

Load stress, tensile bounds, mechanical systems.

17 · banker
Financial Auditor

Ledgers, ROE curves, capital stress models.

18 · analyst
Quantitative Analyst

Regression curves, statistical trends, benchmarks.

Enigma

Enigma is Starship No. 27's proprietary sovereign sparse Mixture-of-Experts — the flagship, closed weights, trained end-to-end on our own stack. Enigma-Gate top-2 dispatch: two experts active per token, load-balanced across a shared attention backbone and eight specialized lobes. Parameter counts and weight footprints stay unpublished.

Ownership
Closed Weights
Not open-weight. Not a rebadge. Not a finetune of anyone's weights — built ground-up by Starship No. 27.
Architecture
Sparse MoE
Enigma-Gate top-2 routing across specialized domain and cognitive lobes.
Deployment
Own Metal
Quantized and served on Starship-operated nodes — no rented third-party model APIs in the path.
Trained End-to-End on Our Stack
Tokenizer
Own byte-level BPE tokenizer, trained from scratch on our own stratified sample.
Corpus
Own multi-source corpus gathered across code, web, math and olympiad reasoning, and distilled SFT data.
Pipeline
Megatron-style sharding with exact dedup and a held-out evaluation split.
Experts on M4
Expert lobes pre-trained on our Apple M4 GPU, assembled into the MoE, and the router trained with load-balancing loss.
H100 Pre-Train
Warm-started into full pre-training on our AWS H100 fleet — ~1.5–2T tokens end to end.
Post-Train
SFT + DPO alignment for instruction following, coding, and reasoning quality.
Serving
Quantized and served on our own nodes — no rented third-party model APIs in the path.
Open Weights
A smaller model from the same family ships open-weight for the community. The flagship stays closed.

Accelerated Intelligence

NVIDIA Inception Member

Starship No. 27 is a member of NVIDIA Inception — access to GPU-accelerated tooling and architectures that help scale sovereign MoE infrastructure at the edge.