Zeraix

LLM inference that fits your hardware. AI tools built around your work.

We build open-source software to make LLMs more useful on the hardware people already own — from the inference runtime to the everyday workspace.

Website · GitHub · X / Twitter

Imparo — the inference engine

Imparo is a hardware- and workload-adaptive LLM inference engine built in Rust.

Less execution overhead: measured tuning and workload-aware execution, with Metal megakernel decode on supported paths.
Reusable context: matching prefixes and shared paged state reduce repeated prefill work across conversations and agents.
Persistent state: retained KV and recurrent state support resuming and branching long conversations and tool loops.
Measured improvements: performance work is paired with output and state-transition checks.

Apple Silicon with Metal is the primary development and validation path. Other backend/model combinations remain qualification-specific. Context reuse does not imply concurrent decoding or continuous batching.

Source & Quick Start · Performance · Contribute

Zeraix — the AI workspace

Zeraix brings models, conversations, agents, files, and tools into an open-source desktop workspace, with local inference and optional cloud providers.

The desktop application and Imparo are separate projects. The current desktop release uses its documented llama.cpp-based runtimes; Imparo is not its default bundled engine.

Download Zeraix · Source & Docs

On Hugging Face

Repository	Purpose
imparo-benchmarks	Published inference benchmark results and measurement notes.
llama-builds	Runtime and sandbox assets consumed by the Zeraix application.
llama-seeds	Version-matched prefix-cache seeds and auxiliary drafter assets for Zeraix.

The two application asset repositories keep their existing download paths. They are not general-purpose standalone chat models or Imparo source mirrors.

Recent updates

2026-09-10: Imparo extended validation to Apple M4 Pro, reproducing optimization benefits demonstrated on M3 Pro. The published four-engine numerical table currently covers M3 Pro.
2026-09-09: Zeraix 2.0 released.
2026-09-07: Imparo published Metal megakernel decode and interleaved M3 Pro comparisons with llama.cpp, oMLX, and rapid-mlx.

Build with us

We welcome contributions to model support, kernels, hardware validation, state management, benchmarks, integrations, and documentation. For performance changes, include a reproducible baseline and correctness checks.

Imparo issues · Zeraix issues · Contact

Project source licenses and third-party artifact licenses are documented in their respective repositories. Model weights and upstream runtime components retain their own terms.