Verinova, Blab, Solab, Spook, VeriRAG: the five model families under everything we ship

The past two months of AI news — and the past two posts on this blog — were about one question: how do you label what a model makes? Article 50, watermarks, disclosure panels. Underneath it sits a quieter question that never makes headlines and decides far more: whose model is it? Most AI products are a thin surface over someone else’s API, which means someone else’s roadmap, someone else’s terms, someone else’s idea of what happens to your data. When we say “Our models. Our gateway. Your AI.”, the first two words are the unusual part. This post is about them: the five model families we train and serve in-house.

Why five families, not one big model

A single model does not make a product. A real product listens, reads documents, looks at images, remembers what your organisation knows, reasons over all of it and answers — sometimes in text, sometimes out loud. Those are different jobs with different constraints: transcription is judged on accuracy and speed, synthesis on latency and naturalness, retrieval on precision, language on reasoning. So the platform under our products is not one model but five families, each trained for its job: Verinova for language, Solab for speech-to-text, Blab for text-to-speech, Spook for vision and OCR, and VeriRAG for embeddings and reranking.

Verinova: the language layer

Verinova is our LLM family — reasoning, generation and chat, trained and served in-house. It is the layer that drafts, summarises, works out what a question means and what a good answer looks like. When a product of ours “writes” or “thinks”, this is the family doing it — on our infrastructure, not rented per call from a frontier lab.

Blab and Solab: the two directions of speech

Speech is two jobs, not one. Blab is text-to-speech: sub-second streaming synthesis and consented voice cloning — the family that gives our platform its voice. Solab is speech-to-text: fast, accurate transcription across languages — the family that gives it ears. You have met both names before as products, and that is deliberate: the family is the engine, the product is the surface we ship on top of it. Blab the studio is where the Blab family faces users, and Calleague’s voice agents speak through the same family; Solab the product turns what the Solab family hears into cited, searchable notes.

Spook: the eyes

Spook is the vision family: OCR, background removal, image generation and visual understanding. It is how a scanned contract stops being a picture of text and becomes text, how a screenshot becomes something a workflow can act on, how an image gets read rather than merely stored. It is the least glamorous family and the one real back-offices lean on hardest.

VeriRAG: the memory

VeriRAG is the retrieval family — embeddings and reranking that make production RAG precise. We wrote a whole post in June about why production RAG lives or dies on retrieval quality: a model’s answer can only be as grounded as the passages you hand it. VeriRAG is that thesis productised — embeddings to find the candidates, reranking to promote the right ones — so the language layer reasons over the right facts instead of reasoning confidently over the wrong ones.

One gateway serves them all

You reach every family the same way you would reach any frontier model: through Qevron, our OpenAI-compatible gateway. One endpoint, one key discipline, one audit trail — and behind it, the five in-house families sitting alongside 43+ external providers, with routing, caching, monitoring and cost analytics in one place. That coexistence is the point. We are not asking you to bet everything on our models: route to ours where they win, to anyone else’s where they don’t. The gateway keeps the score honest, and nothing locks you in — which is a strange thing for a vendor to engineer on purpose, until you remember we also wrote the post about vendor lock-in.

Owned weights change what you can promise

The reason to train in-house is not pride; it is what ownership lets us promise on top. Because the families are ours, the whole stack can run in the cloud, on-premises or fully isolated — where your data lives is a deployment choice, not a negotiation with five vendors. And our last two posts documented exactly what happens when the layer you depend on belongs to someone else: terms that drift, features that vanish overnight, policies that decide what happens to your audio without asking you. A rented model can do all of that to your product. An owned one cannot — its roadmap, its pricing and its data path answer to the same platform that answers to you.

A product can rent intelligence. A platform has to own it.

This is also where the sovereignty story closes. For teams under Türkiye’s KVKK, the practical consequence of five owned families behind one owned gateway is that the default path for your audio, documents and prompts leads to our models on our infrastructure — not to a third-party API in a third country. When you do route outward to an external provider, that is your choice, made explicitly and visible per call at the gateway. Arpanet Bilişim A.Ş. engineered the platform for the KVKK from its first line, and the five families are why that promise reaches all the way down to the weights. Contact us and we will scope it with you.