Starship No. 27 // N27-1 LLM

The Sovereign
70B MoE Model

N27-1 is a proprietary 70-billion parameter Mixture-of-Experts large language model with 27 specialized experts, trained from scratch on a custom multi-node hardware cluster using Apple MLX. Zero third-party weights. Zero cloud dependency. Designed for deployment across the entire Starship No. 27 product ecosystem via the LYRA API.

Dev Q3 2026
70B Params
27 Experts
NVIDIA H100 Target
LYRA API Q3 2027
70B
Total Parameters
27
Specialized Experts
~$28K
Cluster Budget
2
Node Cluster (M4 + AMD)
Technical Architecture
How N27-1 Is Built
Dev
70B MoE — 27 Experts
Mixture-of-Experts architecture with 27 specialized expert networks. Dynamic routing activates only relevant experts per token. Each expert handles a distinct capability domain.
Dev
Apple MLX Native Training
Primary training node runs on Apple M4 Max via MLX framework — unified memory, Metal compute, zero PyTorch overhead. Secondary AMD mesh node for distributed data-parallel scaling.
Active
NVIDIA H100/A100 Cloud Burst
Pre-training compute unlocked via NVIDIA Inception member access. Utilizing NeMo framework for MoE architecture design and TensorRT-LLM for production optimization.
Planned
Production Inference Stack
TensorRT-LLM compilation for 4x throughput. NIM containers. KV cache optimization, speculative decoding, continuous batching. Target: <50ms TTFT at 70B scale.
New
Domain Fine-Tuning (Q4 2026)
Three specialist heads: Music (lyrics, composition, production), Legal (contracts, IP, licensing), Tech (code, architecture, systems). Curated proprietary datasets per domain.
New
RLHF + Alignment (Q1 2027)
Constitutional AI training with N27-specific values: sovereignty, creator-first economics, anti-surveillance, transparent reasoning. Human feedback from operator team.
Planned
LYRA API Integration
Native deployment via LYRA inference router. Governor rate limits, token budgets, cascade fallback (N27-1, MikuOS, Gemini). SysOp cockpit telemetry.
Planned
On-Chain Model Registry
LyraLedger anchor for model checkpoints, training runs, and benchmark scores. Immutable provenance. LYRA token gates API tiers.
Training Pipeline Stages
Stage 1 — Pre-Training (Q3 2026)
Curated multi-domain corpus (~2T tokens). Music, legal, tech, code, scientific literature, multilingual. MLX distributed data-parallel across M4 primary + AMD secondary. Status: Unblocked via NVIDIA Inception cloud burst credits.
Stage 2 — Domain Fine-Tuning (Q4 2026)
Three parallel LoRA/QLoRA adapters: Music, Legal, Tech. Evaluation via domain-specific benchmarks + operator blind tests.
Stage 3 — RLHF + Alignment (Q1 2027)
Constitutional AI with N27 values. Preference data from SysOp team + artist partners. Reward model trained on LYRA infrastructure.
Stage 4 — Production Deployment (Q3 2027)
TensorRT-LLM optimized engine. NIM microservice containers. LYRA API router with governor budgets. CYPH chat integration. Orchestrator agent access.
Roadmap
Milestones to Sovereign Intelligence
Dev
Cluster Bring-Up + MLX Pipeline
M4 primary + AMD mesh node operational. MLX distributed training pipeline validated end-to-end.
Inception
DGX Cloud H100 Access
NVIDIA Inception member access unlocked. Provisioning high-throughput GPU infrastructure and cloud credits for 70B compute pre-training.
Dev
Pre-Training Run (70B, 27 Experts)
Full converged run on curated 2T token corpus. Checkpoint every 50B tokens. Target: <2.5 perplexity.
New
Music / Legal / Tech Fine-Tunes
Three specialist LoRA adapters trained in parallel. Domain evals: lyric quality, contract accuracy, code pass@k.
Next
RLHF + Constitutional Alignment
Preference data collection. Reward model training. PPO/DPO runs. Red-team evaluation by Agent 6/8.
Next
TensorRT-LLM + NIM Production
Engine compilation, quantization (FP8/INT4), KV cache tuning, speculative decoding. Load test at 100+ concurrent.
Future
LYRA API General Availability
N27-1 live on LYRA router. Governor budgets per tier. CYPH chat integration. Orchestrator agent access.
Future
N27-2 Architecture Research
Post-70B scaling: 200B+ params, 64+ experts, multimodal (audio + text). Next-gen cluster: 8x H100 or GB200 NVL72.
Technology
Core Stack

Training Stack

  • Apple MLX (primary)
  • PyTorch (AMD node / eval)
  • NeMo Framework (MoE arch)
  • Tokenizers (Rust)
  • HuggingFace Hub (checkpoints)

Production Stack

  • TensorRT-LLM
  • NVIDIA NIM Microservices
  • Triton Inference Server
  • Kubernetes (KServe)
  • Prometheus + Grafana

Hardware Targets

  • Apple M4 Max (MLX dev)
  • AMD RDNA 3 (mesh worker)
  • NVIDIA H100/A100 (pre-train)
  • NVIDIA GB200 NVL72 (future)

Integration Surface

  • LYRA API Router
  • Governor Rate Limits
  • CYPH Chat
  • Orchestrator Agents 1-7
  • VanguardOS Think Module
  • MikuOS Creative Cores

Accelerated Intelligence

Member of NVIDIA Inception

Starship No. 27 is a member of the NVIDIA Inception program. This program provides critical access to NVIDIA's industry-leading GPU-accelerated computing tools, technical expertise, and custom hardware architectures, accelerating our mission to pre-train, compile, and scale sovereign mixture-of-experts model infrastructure at the edge.

© 2024 NVIDIA Corporation. NVIDIA, the NVIDIA logo, and NVIDIA Inception are trademarks and/or registered trademarks of NVIDIA Corporation in the U.S. and other countries.

One model.
Zero dependencies.

Every major AI lab rents compute from the same three cloud providers. N27-1 breaks that chain — trained on hardware we own, deployed on infrastructure we control, serving products we built. No API keys from OpenAI, Anthropic, or Google. The model belongs to the network.

LYRA Platform Nullxre