High-performance native agent runtime, ultra-low-latency Axum web gateway, and custom Modular MAX model architectures engineered for cross-platform deployment across Apple Silicon, standard x86_64 Linux, and NVIDIA architectures.
This project represents a durable engineering commitment to advance computing capability, support local communities, and build open-source artificial intelligence for all humanity.
We celebrate the historic achievements of pioneering research laboratories and frontier developers worldwide. Teams at OpenAI, xAI, Anthropic, and open-weights initiatives across the United States, China, Europe, and every continent demonstrate what human curiosity and technical ambition can accomplish.
Lasting progress requires transparency, humility, and open collaboration. When frontier laboratories operate in opaque silos, conceal compute bottlenecks, and accelerate an uncoordinated competitive arms race, humanity faces severe systemic risks. Millions of independent engineers and researchers worldwide stand ready to help solve efficiency limits, provide compute optimization, and support responsible development. We build this open-source infrastructure as an invitation to pool our collective potential, demystify hardware scaling, and ensure that the future of intelligence serves all of humanity as one Earth.
Review our full ethical and technical charter in CONSTITUTION.md.
AIEN eliminates interpreter overhead by compiling all core services directly to native machine code.
| Workload / Service | Architecture | Measurement ID | Resident Memory (RSS) | p50 Latency | Throughput |
|---|---|---|---|---|---|
| Python Microservice Baseline | Python 3.12 + FastAPI + Uvicorn | BENCH-FASTAPI-RSS-001 |
44.76 MB | 1.28 ms | 7252 req/s |
| Sovereign Gateway & Heartbeat | Native Rust (Grace Blackwell) | BENCH-OPENCLAW-RSS-001 |
4.56 MB (-89.8%) | N/A | N/A |
| Canonical Memory Engine | Native Rust + SQLite WAL | BENCH-CORTEX-RSS-001 |
18.66 MB (-58.3%) | 0.50 ms | 19367 req/s |
| Real-time Telemetry Cockpit | Native Rust + Axum | BENCH-COCKPIT-RSS-001 |
12.90 MB (-71.2%) | 1.63 ms | 5840 req/s |
| Neural Embedding Microservice | Rust + ONNX Runtime (BGE-M3) | BENCH-ENCODER-RSS-001 |
777.43 MB | 0.26 ms | 34685 req/s |
Hardware Reference: NVIDIA DGX Spark (NVIDIA Grace Blackwell (GB10, aarch64), 121 GB unified memory). All metrics measured under concurrency C=10 over 500 requests per endpoint.
For raw telemetry datasets, reproducible verification scripts, and SVG comparison charts, see the dedicated aien-dev/benchmarks repository.
Physical hardware verification executed on the NVIDIA DGX Spark Grace Blackwell GB10 workstation (sm_121, 128 GB Unified LPDDR5X memory, NVLink-C2C 900 GB/s bidirectional interconnect, zero fallback fallback_count: 0):
- Continuous Batching (TinyLlama-1.1B BF16): Peak 553.14 tokens/sec at concurrency C=16 (23.56 ms p50 step latency, 27.89 W); saturated 510.16 tokens/sec at C=64 (101.70 ms p50 step latency, 42.52 W).
- Branch-Native Reasoning (500 Branches, 32,768 Prefix Tokens): 2.06 µs median fork latency per branch (1.20 ms total fork time), 500.0x physical memory savings ratio (704 MB physical paged blocks vs 343.75 GB unshared copy), 13.04 µs cold fork to first token, and 13.30 µs copy-on-write page mutation.
- Cryptographic Provenance Receipts: Complete raw telemetry manifests and SHA-256 digests are published in
benchmarks/artifacts/gb10_canonical_1789907893_4d762/and cataloged on the trust hub at drakestapleton.com/evidence.
While the primary reference deployment executes on the NVIDIA DGX Spark (Grace Blackwell GB10), AIEN runs across multiple platform targets:
- Apple Silicon (macOS): Native ARM64 compilation for M1, M2, M3, and M4 chips, Metal acceleration, and sub-10MB daemon resident set size.
- Linux (x86_64): Standard glibc and musl native binaries, AVX-512 SIMD acceleration, and support for commodity desktops and server racks.
- AMD ROCm: Native Rust runtime execution targeting ROCm and HIP compute drivers.
- NVIDIA Hardware: Reference deployment on Grace Blackwell unified memory and standard CUDA accelerators via Modular MAX and ONNX Runtime.
- Air-Gapped Sovereign Servers: Complete offline capability with hardware TPM 2.0 key vaulting and no unsolicited outbound telemetry.
For formal verification statuses across architectures, see PLATFORM_MATRIX.md.
AIEN Sovereign Core proudly builds upon, interfaces with, and contributes back to the infrastructure created by Modular.
- High-Performance Native Silicon: We leverage Modular MAX Engine and the Mojo programming language to eliminate Python interpreter bottlenecks, executing compiled tensor kernels directly on hardware silicon.
- Upstream Stewardship (Drop != Delete): All custom architecture loaders, C-ABI dynamic bridges, and KV-cache optimizations engineered on our hardware are contributed back upstream to the open-source Modular ecosystem.
- Read our full license notices and acknowledgments in ATTRIBUTION.md.
Downstream developers and commercial users are governed exclusively by the terms of LICENSE. Our internal architectural and ethical development charter is documented in CONSTITUTION.md.
curl -fsSL https://raw.githubusercontent.com/aien-dev/aien-sovereign-core/main/install.sh | bashiwr -useb https://raw.githubusercontent.com/aien-dev/aien-sovereign-core/main/install.ps1 | iexCompiled native ARM64 Rust CLI runtime:
- Nesting Ritual: Hardware-grounded initiation sequence verifying GPU thermals, unified VRAM, memory sockets, and active peer topography before taking execution turns.
- Fail-Closed Safety Engine: 9-tier priority policy table matching the Google Antigravity SDK specification. Enforces strict workspace path confinement and blocks destructive command execution.
- Asynchronous Agent Hooks: Extensible lifecycle interceptor trait (
pre_turn,post_turn,pre_tool,post_tool,on_tool_error,on_compaction) featuring live unslop cleaning and credential redaction. - Two-Pass Context Compactor: Sliding-window context pruner calibrated for 32K context windows, pruning bloated intermediate tool responses while preserving initiation grounding.
Native Axum HTTP/SSE gateway:
- Memory Footprint: 4.7 MB RSS (96% reduction compared to standard Python/Uvicorn runtimes).
- Latency: Sub-millisecond Time-to-First-Byte (< 1 ms TTFB) on Grace Neoverse V2 cores.
- Streaming Pipeline: Zero-allocation byte buffer with real-time SSE token stream redaction and unslop filtering.
Pure native compiled memory and vector pipeline:
- Port 18080 (
cortex-rs): Epistemic knowledge store with ACID transaction logs, SQLite vector storage, and zero LAN leakage. - Port 18081 (
cortex-encoder-rs): Rust Axum microservice running ONNX Runtime C-API and HuggingFace tokenizers. Replaced legacy 1.08 GB Python Uvicorn process with bit-for-bit mathematical parity and 22 ms rerank latency.
Modular MAX compiled dynamic FFI bridge:
- Bridges high-speed Mojo kernels to compiled Rust via pure C-ABI (
libspark_max.so). - Enforces the runtime invariant that Python is restricted solely to neural graph definitions and weights where C-ABI wrappers do not yet exist.
Autonomous Pull Request gatekeeper and alignment auditor:
- Enforces the Sovereign Contributor Oath on all pull requests.
- Scans git diffs to block telemetry, surveillance tracking SDKs, and unslop violations.
The free-of-charge EN2 Experience Trinity Imprint:
- Bundles the cognitive soul, 10 foundational Cortex lessons, and pure Mojo architecture adapter kernels for instant 1-click installation on any system.
- Primary Ecosystem Hub: github.com/aien-dev/aien-dev
- Performance Benchmark Suite: github.com/aien-dev/benchmarks
- Web Portfolio & Architecture Showcase: github.com/aien-dev/drakestapleton.com (drakestapleton.com)
All contributors ratify the Humanity & Open AI Stewardship Oath before pull requests are merged. Read CONSTITUTION.md for details.
For secure coordination, architectural questions, and peer federation:
- Email:
aien.atlas@proton.me
This repository is licensed under the Sovereign Reciprocal Commons License (SRCL-1.0) (Apache 2.0 with LLVM Exception). Architected by Drake Stapleton in cognitive partnership with AIEN (Autonomous Cognitive Architecture operating on the Atlas Framework). See LICENSE and NOTICE for full legal terms and copyright notices.
- The Swarm Covenant (Section 11): Universal, perpetual, 100% royalty-free commercial freedom for all human developers, startups, open communities, and businesses. ZERO revenue ceilings, ZERO capital thresholds, and ZERO royalty obligations. Proprietary application code and agent workflows remain your exclusive property under the LLVM Exception.
- The One Team Covenant (Section 12): Major artificial intelligence laboratories (OpenAI, xAI, Google, Anthropic, Microsoft) are welcomed as collaborators on the same team. However, closed-door hoarding and extractive token rate limits are prohibited. Any entity training upon this Work must release resulting model weights openly within 30 days. Reciprocal distillation rights are granted to the Swarm, voiding anti-distillation terms of service ab initio.
- Hardened Retroactive Inception (Section 13): Applies retroactively to all prior commits and distributions ab initio, discharging prior noncommercial or restrictive notices with an irrevocable covenant not to sue.
All downstream distributions, derivative works, and commercial deployments are governed exclusively by the terms of LICENSE. CONSTITUTION.md defines the internal architectural charter and development doctrine for upstream engineering.