Salesteq is a member of NVIDIA Inception. Here is what actually runs on NVIDIA GPUs in our stack today: our Arabic speech models, and a small dealer-specialised model that answers in about a second where a hosted frontier model takes twelve. And why, for a dealership, latency and a fixed cost line matter more than the biggest model.
Salesteq is a member of NVIDIA Inception, NVIDIA's programme for startups building on its platform. This post is the plain version of what that means for us: what runs on NVIDIA in our stack today, what is still research, and why a dealer should care about any of it.
Salesteq is the AI platform for automotive retail. It runs a dealer's or importer's front office and back office: customers answered on WhatsApp, Instagram, the website and the phone, in their language, from live stock and list prices; service appointments written into the dealer management system with the slot; quiet leads followed up; every call transcribed and scored. It is live at BMW, MINI, Ford and Hyundai retailers in Saudi Arabia, at a nine-brand marketplace there, and at dealers in Switzerland and Tajikistan.
Most of what a customer sees is the work of an agent calling tools against the dealer's own systems. Two layers under that are models we train and serve ourselves, on NVIDIA GPUs. Those two layers are the reason we are in Inception.
What runs on NVIDIA today
Our speech models. Real dealer calls break general-purpose speech engines: Gulf dialects, car model names, part numbers, two languages in one sentence, a workshop in the background. So we build our own. Our Arabic speech recognition took first place at NADI 2026 for mixed-dialect Arabic, and on real BMW service calls it reaches a 17.4% word error rate where the three engines we measured against sit at 21.5%, 30.8% and 31.4%. These models are trained and served on our own NVIDIA cards in Germany, with CUDA and cuDNN underneath. Nothing about a customer's voice leaves hardware we control.
A contribution back. Along the way our team translated C++ speech libraries to Rust for the ecosystem. A two-minute text-to-speech job that took twenty seconds now takes five.
The inference story: small, specialised, on one card
The second layer is the one we are most pleased about, and it started as a question rather than a plan.
The language model behind our agents is, in production, a hosted frontier model. It is good. It is also slow and it gets more expensive with every conversation a dealer wins. So our AI lead, Osama Abdelaal, ran an experiment: take a small open-weight model (Qwen3, in the 4B to 14B range), give it our prompt and our tools, and put it through the same dealer work as the hosted model. Live stock, pricing, finance, booking into the dealer management system.
The results, on a single NVIDIA card:
- 82% against 63% on our dealer-work suite. The small model, specialised only by our prompt and tools, beat the hosted frontier model on the work that pays. To be fair to the frontier model: hand both the identical prompt on a controlled run and they tie at 84% each. A small owned model holds the frontier API to a draw, and wins where it has been shaped to the operator.
- About one second a turn on the card, around twelve seconds a turn hosted. The hosted model's own reasoning is the latency, not our tools, and it does not shrink with scale. There is no on-device option for it at all.
- The system wins, not the model. A general model with no tools scores about half and invents cars. Grounding and tools are what make the answers true, and we own those.
Two sizes matter in practice. A 4-bit Qwen3-4B fits in about 8 GB and answers in roughly 880 ms, within two points of the 14B: that is a Jetson-class box in a dealership. An RTX PRO 6000 Blackwell serves the 14B at full quality for a central tier. We have run the whole agent loop on-device, grounded in live inventory, with nothing leaving the machine.
Underneath sits a serving runtime Osama wrote for exactly this: 853 lines where a stock serving library has around 52,000, inside a microVM at about 2% overhead, with no direct host system calls and no network listeners, delivering 90% of the reference runtime's throughput at 1.1B parameters and 97% at 13B, with byte-identical output. Measured on NVIDIA GPUs on the non-confidential tier.
Why latency and cost, not model size
Three reasons, in the order a dealer would put them.
Latency is the product. A customer on the phone, or on WhatsApp, waits for an answer. One second feels like a person. Twelve seconds feels like a form. Nothing else we do about tone or wording recovers those eleven seconds.
A cost line that stops growing. One tenant's hosted inference costs us about $488 a month today, and that number grows with every conversation. On an owned card the same work costs 14 to 197 times less, depending on load, and breaks even at three or four dealers. A GCC dealer group already spends $500,000 to $1,000,000 a year on its call-centre function. Inference should be a fixed asset in that budget, not a per-token tax on growth.
Sovereignty becomes a product. When an importer's regulator asks where the models run, "on a card in the building" is the answer that passes. A box you plug in, and the dealership runs.
The honest boundary
The self-hosted base model is proven: it runs the full agent loop on one NVIDIA card today. The automotive fine-tune on top is still research. Our training pipeline runs end to end on our own machines, the first run did not beat the base, and closing that gap is data and a stable GPU, not a breakthrough. We will publish the numbers when it is done, not before.
The confidential-compute measurements for the runtime need hardware we do not yet own. That is part of why Inception matters to us: access to Blackwell-class capacity, and to the people at NVIDIA who have solved these problems before.
What comes next
Speech stays ours and stays on NVIDIA. The next step is bringing LLM inference in-house on an RTX PRO 6000 Blackwell node, for latency, fixed cost and jurisdiction control, with the dealership box as the destination.
If you run a dealer group or an importer and the numbers above look like your call centre, the live platform is at salesteq.com. If you work on inference at NVIDIA and want to see an agent write a booking into a real dealer system in a second, write to us.
Numbers in this post: speech figures from our published Arabic speech recognition evaluation; model comparison from our dealer-work suite, July 2026; platform figures as counted on production in October 2026 (over 15,000 workshop bookings written into dealer systems since April, 48% of customer conversations outside 08:00 to 18:00).
