Local-first
Run inference locally and retain control over hardware, latency, privacy, models, and execution.
Public R&D Release
AIR 0.9.12
AIR
Adaptive Inference Runtime
A local-first inference runtime built around explicit architecture, observable runtime state, bounded execution, and real qualification.
AIR is an open Superior MI Labs research effort exploring how inference systems can become more transparent, measurable, adaptable, and understandable.
AIR
runtime visible
LOCAL-FIRST
OBSERVABLE
BOUNDED
Runtime principles
Intelligence is easier to trust when its execution is visible.
AIR treats inference as an engineered runtime rather than an invisible call into a black box. State, decisions, execution boundaries, and qualification should remain inspectable.
Observable state
Expose what the runtime is doing instead of hiding important behavior behind opaque abstractions.
Explicit architecture
Clear contracts and bounded components make the system easier to reason about, test, and evolve.
Qualification first
Experimental behavior should survive evidence, measurement, comparison, and falsification.
Architecture
Follow the flow.
AIR exposes the major stages of runtime behavior as distinct, inspectable surfaces instead of collapsing everything into one hidden inference path.
01
Request
explicit input
→
02
Runtime
bounded execution
→
03
Inference
model execution
→
04
Decision
measurable scoring
→
05
State
observable result
Hover across the runtime stages. The background activity responds to the same idea: signals travel through a visible system instead of disappearing into an opaque execution path.
Research surface
Build. Measure. Falsify. Keep what survives.
AIR is developed as an R&D system. Experimental ideas are useful only when they can be implemented, observed, challenged, and compared against alternatives.
A
Generation
Model execution remains one part of a larger runtime, rather than the architecture itself.
B
Semantic decision
Candidate evaluation can be surfaced explicitly, measured, and tested independently.
C
Runtime visibility
Diagnostics, state, metrics, configuration, and execution should remain accessible to researchers and operators.
D
Hardware-aware evolution
AIR is designed to evolve through measured interaction between software architecture and the machines executing it.
Public release
AIR 0.9.12
The current public R&D release exposes AIR's engineering direction, runtime interface, and experimental decision surface.
Runtime
Local-first
Architecture
Explicit
State
Observable
Research
Open
Release
0.9.12
AIR is not trying to hide complexity. It is trying to make complexity understandable.
GitHub remains the canonical engineering repository. Hugging Face is the public-facing research and demonstration surface.
Inspect the source and release history ↗Superior MI Labs
Local roots. Open research. Wider horizons.
Built in Upper Michigan. Designed to be inspected, tested, and improved in the open.
