The Universal AI Inference Engine

The artificial-intelligence inference stack is fragmented and locked in. Today, running different classes of models in production means assembling and maintaining a different software stack for each one — one engine for language models, another set of glue code for image and video diffusion, yet another for speech — and almost all of these tools are tied, in practice, to a single hardware vendor's proprietary software. For European public bodies and regulated organisations, this fragmentation compounds a deeper concern: a structural dependence on non-European technology for one of the most strategic layers of the modern digital economy.

NeuroBrix is a universal AI inference engine that removes this fragmentation. It executes any model — large language models, vision-language models, image and video diffusion, speech recognition and synthesis — through a single runtime, a single portable model format, and a single hardware-placement solver, with no model-specific code. It is open-core (Apache 2.0), designed from the ground up to be hardware-agnostic, and engineered for auditability and on-premise sovereignty rather than dependence on any external cloud or vendor account.

This document sets out the problem NeuroBrix addresses, its technical foundations, the evidence base behind its claims, an honest comparison with existing inference engines, and its alignment with Europe's emerging technological-sovereignty agenda. NeuroBrix is designed and developed by Hocine Benkelaya in Alicante, Spain, and distributed commercially under licence by WizWorks OÜ, an Estonian company: the engineering and the commercial operation of the project both sit inside the European Union.

Modern artificial intelligence is no longer a single kind of model. A production system today may combine a large language model for reasoning, a vision-language model for document understanding, a diffusion model for image or video generation, and speech models for transcription and synthesis. Each of these has, in the open-source ecosystem, grown its own serving stack.

The result is structural fragmentation. Language models are served by one family of engines; image and video diffusion are run through a different set of pipelines and user interfaces; speech requires yet other tooling. Each stack carries its own dependencies, its own deployment model, its own failure modes and its own operational expertise. An organisation that wants to deploy several modalities must integrate, secure, patch and staff several unrelated systems at once. There is, today, no unified, industrial-grade runtime that treats all of these the same way.

For European institutions, the stakes are explicit. On 3 June 2026 the European Commission presented its European Technological Sovereignty Package, warning that over-reliance on non-EU providers for cloud and AI infrastructure poses a significant risk to Europe's digital autonomy and resilience, and positioning open source as a strategic instrument to reduce vendor lock-in, improve transparency, and build European alternatives across the technology stack. The accompanying Cloud and AI Development Act introduces a sovereignty-assurance framework in which control over the software supply chain, transparency, and infrastructure location become explicit criteria for public-sector and critical-infrastructure use.

An inference layer that is fragmented by modality, dependent on a single foreign hardware ecosystem, and opaque in its software supply chain is precisely the kind of dependency this agenda seeks to reduce. The need is for an inference engine that is unified across modalities, portable across hardware, auditable in its dependencies, and deployable entirely on-premise under the operator's own control.

NeuroBrix answers that description on grounds that can be checked rather than asserted: it is developed in Alicante, Spain, and managed from Estonia, so both the engineering and the operation sit inside the Union; its engine is Apache 2.0, readable and forkable; it runs on the operator's own machines with no account, no external registry and no outbound call; and its dependency chain at inference is small enough to be enumerated and is verified by the operator's own probe. What follows argues from those properties, not from a flag.

NeuroBrix is built on a single founding premise: every model is, at bottom, a graph of mathematical operations executed on hardware. A language model, a diffusion transformer, a speech recogniser and a video generator differ in architecture, but each reduces to a computation graph of standard tensor operations. NeuroBrix is the engine that takes that graph and computes it — efficiently, on whatever hardware is available, regardless of which modality the model belongs to.

Concretely, NeuroBrix replaces a fragmented collection of per-modality stacks with three unified components:

Several tools are excellent at one thing. llama.cpp runs language models across an exceptionally wide range of hardware, but only language and vision-language models, in a single specialised format. vLLM is the reference for high-throughput language-model serving and has extended toward multimodality through a separate companion project that stitches together many third-party libraries beneath it. NVIDIA NIM delivers highly optimised inference but is bound to one hardware vendor and, in its standard mode, to that vendor's online account and registry. ONNX Runtime is a portable graph executor with many hardware back-ends, but it is not a multimodal serving runtime: the multimodal pipelines, schedulers and memory management must be assembled around it.

NeuroBrix's distinctive position is the combination, in one system, of four properties that no competitor offers together: all modalities under one runtime; hardware-vendor-neutral by design; zero third-party model dependencies at inference; and reproducible, auditable artifacts — delivered as an open-core engine.

NeuroBrix follows an open-core model. The inference engine — the runtime that executes .nbx artifacts — is fully open source under the Apache 2.0 licence, freely available on public repositories and the Python package index, and free to adopt, inspect and self-host. Around this open core sits a commercial layer: the proprietary toolchain that converts and prepares models, and the model hub that distributes validated artifacts. This is a well-established structure for deep-technology infrastructure companies: an open, trustable, adoptable core, with sustainable commercial tooling and distribution on top.

The open core is deliberately the part that matters most for sovereignty and trust. An organisation can read the engine's source, verify what it does, and run it without depending on the project's commercial services. Licensing is clean: the engine carries no model licence of its own; each model retains the licence of its original author, recorded transparently in the hub.

The public presence is live today: the engine is distributed through public source repositories and the Python package index, and the model hub (neurobrix.org) serves validated artifacts with per-model licence metadata.

NeuroBrix is designed and developed by Hocine Benkelaya in Alicante, Spain, on a privately operated GPU cluster, and NeuroBrix and its intellectual property belong to Hocine Benkelaya personally. The project is supported and funded by Neural Network Holding Ltd, a company incorporated in the United Kingdom, and distributed commercially under licence by WizWorks OÜ, a company incorporated in Estonia.

Development therefore takes place in Spain and operations are run from Estonia — both within the European Union — under United Kingdom ownership of the operating company. This document announces no other domiciliation: where the project may be established in future is not decided, and an undecided thing is not a claim.

The engine is at public release 0.6.4 (Apache 2.0). Nine model families are validated end-to-end through the four-mode method; 49 models are published on the public hub; and the dependency-free property and build reproducibility hold across the validated set. Empirically, validation to date has been performed on NVIDIA Volta-class accelerators and on Apple Silicon. The architecture is vendor-neutral and the additional accelerator paths are engineered work, not a research question.

One of those paths is further along than the others and is reported here as work in progress rather than as a delivered feature. Triton execution on Apple GPUs is proven on a public development branch: first light is passed, a complete language model has been executed end to end, and the memory allocator and the kernel launcher are written. It is not on the main branch and it is not in the release above; it is evidence that the vendor-neutral design holds outside the platform it was built on, and nothing more than that yet.

The inference layer of the AI stack is where models meet hardware — and it is today both fragmented by modality and concentrated on a single foreign hardware ecosystem. NeuroBrix offers a different foundation: one open-core engine that runs any model of any modality, on a hardware-agnostic basis, with an auditable and dependency-free inference path and reproducible artifacts. It is developed in Alicante, Spain, and managed from Estonia — engineering and operation both inside the Union — aligned by construction with Europe's technological-sovereignty objectives, and validated across nine model families, with 49 models published on the public hub.

We welcome the opportunity to demonstrate the engine and to explore where a sovereign, universal inference engine developed in Spain can contribute across the national and European AI ecosystem. Technical teams, deployment operators and compute partners are invited to engage with the project directly — to evaluate the engine on their own models and hardware, and to discuss concrete deployments.