Execution must remain stable
under real-world conditions.
Modern AI systems rarely fail because models are mathematically wrong. They fail because execution assumptions break — shared GPUs, transient memory pressure, scheduler jitter, models that don’t fit where you expect them to.
Atenia Engine is an LLM inference runtime written in Rust from scratch. It treats execution as a first-class concern: a system layer that can observe, reason, and adapt without ever touching what the model computes.
Today it runs seven model families — Llama, Qwen, Gemma, Phi, Mistral, SmolLM and Falcon3 — from a single command-line tool, including an interactive chat. It runs Llama 2 13B Chat on a laptop with 8 GB of VRAM and 32 GB of RAM, spreading weights automatically across VRAM, RAM and NVMe. Every certified output is validated against F64 mathematical ground truth.
- Stability first Peak throughput means nothing if execution collapses in production. Performance optimization is done after correctness is locked — not before.
- No semantic drift Adaptation happens strictly at the execution level. The model’s numerical output is never modified by execution decisions. What the model computes is always what the model computes.
- Explicit decisions Every execution decision is bounded and observable. There are no opaque heuristics. Where the engine places a tensor, why it spills, and what it restores are all traceable.
-
Behavior over claims
Core properties are validated through executable tests, not benchmarks.
If it isn’t backed by
cargo test, it isn’t documented as complete. - Reality-first Designed for noisy, shared, imperfect hardware. The beyond-VRAM demo runs on a consumer laptop, not a data center. That’s the point.
One binary. One command.
Your own model.
Atenia is a single Rust binary with no Python environment to manage and no server to keep running. Install it, point it at a HuggingFace or GGUF checkpoint, and start generating or chatting.
The command line is designed to be friendly: atenia doctor
checks your machine, atenia diagnose checks a model before
you load it, atenia chat opens an interactive conversation,
and every error message explains what happened and how to fix it.
Atenia Engine does not ask
to be trusted blindly.
Every capability documented on this site is backed by an executable test
in the public repository. The F64 ground-truth validation numbers are
reproducible with cargo test against real model weights.
The beyond-VRAM demo runs in approximately 7 minutes on a 32 GB / 8 GB box.
This is why the project places unusual emphasis on explicit non-goals, published ADRs (Architectural Decision Records), and milestone handoffs that document what was built, what failed, and why decisions changed.
Execution intelligence should earn trust through observable behavior — not through marketing claims.
📄 Paper — Atenia Engine: Hardware-Adaptive Execution Intelligence for Stable and Resilient AI Runtime Systems. Preprint withdrawn pending maturity of the underlying engine. A revised write-up with full empirical validation will follow.
🧾 Patent — USPTO Provisional Application No. 63/941,875. Filed December 16, 2025.
These are deliberate boundaries, not missing features.
- A machine learning framework
- A compiler or graph optimizer
- A heuristic tuning layer
- A performance-at-all-costs system
- An opaque, black-box adaptation mechanism
- A replacement for PyTorch, JAX, or similar
Atenia complements existing frameworks by addressing the layer they largely ignore: execution stability under real-world conditions.
- Repository github.com/AteniaEngine/ateniaengine
- Roadmap Current milestones and next directions →
-
Tests
Behavior-driven validation —
cargo test