← Home
// Product family

AI infrastructure

Our own inference engines, to run AI locally. The open-source EIE engine and the measurement protocol that accompanies it.

EIE · Elyne Inference EngineOpen Source · Apache 2.0

A local GGUF inference engine, from desktop to mobile.

EIE serves chat and embeddings over HTTP. One model is enough; multiple models can be configured for parallel, sequential or fan-out execution. Application memory, documents and tools remain the responsibility of clients.

  • Partial OpenAI-format API compatibility, with SSE streaming.
  • Windows x64 CPU, macOS and Android arm64 bundles available.
  • Local Windows/CUDA, Windows CPU, Intel macOS and Android CPU tests; Metal use reported by the maintainer.
  • EWS: experimental expert-weight streaming for selected MoE models.
› Technical details
C++17 server using identified llama.cpp runtimes. Groups are implemented without full load qualification; automatic KV-cache adaptation is not operational. A Gemma 26B EWS pilot measured CUDA weight buffers of 14.424 → 4.789 decimal GB at 32 slots per layer: this is not total VRAM. A separate GLM-5.3-Flash 320B runtime completed a Next roundtrip on an RTX 4090 Laptop with 16 GiB-class VRAM and 32 GiB-class RAM: 148 tokens in 278.515 s, with CPU/GPU expert computation. These results establish feasibility in their profiles, not a general speed advantage. Android CPU text bundles were tested on a Z Flip6 / Android 16, with no APK or EWS, vision or NPU qualification. Model weights retain their own licences.
Elyne BenchmarkPublic protocol · Apache 2.0 / CC-BY-4.0

Measuring what a persistent local-AI substrate adds to the same model.

A public three-arm benchmark: the full system, the surfaced information alone and the bare model, with the same model binding, sampling parameters and seed. The repository publishes the instrument, frozen scoring, selected evidence and an adapter protocol; it does not publish Next’s private runtime.

  • Three frozen batteries: accuracy (17 cases), orchestration (14) and tool chain (15).
  • Reference results, system / information / bare model: 15/15 · 14/15 · 6/15, 14/14 · 12/14 · 0/14 and 15/15 · 0/15 · 0/15.
  • Published report verifiable by script and unit tests, with SHA-256 anchors.
› Technical details
Reference observations from a single deterministic campaign, not general performance estimates; two judge-scored accuracy items remain pending. The adapter must attest identical model and sampling identities for every arm, otherwise the harness fails closed. The mock adapter only exercises the protocol and its scores are not results. Software under Apache 2.0; batteries, reports and documentation under CC-BY-4.0.

Let’s discuss your project.

Contact us