BERLIN, October 21, 2026. SPREAD, the AI for Engineering company, today announced at NVIDIA GTC Berlin that it has post-trained an NVIDIA Nemotron 3.5 Lightning 30B-A3B open model using NeMo Automodel and NVIDIA Model Optimizer to create a compact, 4-bit engineering model. JLR is adopting the model to support one of the most time-critical jobs in vehicle development: processing engineering tickets. The luxury vehicle manufacturer is launching five new vehicles over the next 18 months, including the Range Rover Electric, Range Rover Sport Electric and Jaguar Type 01. On 1,200 real-world engineering tasks, the model outperformed leading frontier models and a 4x larger Nemotron model on 9 of 10 measures.
Why it matters
Vehicle launches are getting harder to land. Engineering data sits across PLM, requirements, E/E, and software tools, in schemas no general-purpose model has seen. Meanwhile every launch has to hit its start of production (SOP) and deliver the customer experience its price point promises.
The use case: engineering tickets
During development, validation, and launch, engineers raise thousands of tickets: defects found in testing, change requests, and early field issues. Each ticket needs three answers fast. Which components, functions, and software versions does it touch? Has a similar issue been solved before? Who owns the fix? Today engineers often search across tools by hand to find those answers.
SPREAD’s model reads each ticket in plain language and searches across the ticket history and SPREAD’s Product Twin of the vehicle. It links the ticket to the affected parts of the product and surfaces similar resolved cases, returning a median answer in 0.36 seconds. Engineers spend their time fixing issues instead of looking for context, and fewer issues reach SOP or the customer.
The results
On ticket analysis, the model scored 79% against 44% for the best frontier model. On dataset queries it scored 81% against 52%, and on multi-hop questions that span several systems it roughly doubled the best frontier result. It also produced fully correct data mappings twice as often, which matters for connecting new data sources into the Product Twin.
How SPREAD built it
SPREAD fine-tuned the model on 5,061 examples using NeMo Automodel and compressed it to NVIDIA’s NVFP4 4-bit format using Model Optimizer open library. A short quantization-aware distillation step then recovered the accuracy lost to compression. The full pipeline took 4.3 hours on NVIDIA H200 GPUs.
On a single NVIDIA Blackwell GPU, the model uses 1/35 of the base model’s output tokens, and its fastest variant serves 5 requests per second. That is more cost effective to run on every ticket and small enough to deploy in the customer’s own environment.
Philipp Noll, co-founder and MD of SPREAD, said:
Engineering teams need AI that knows their product logic, and a small model trained on the right data beats a giant generalist on the work that matters, while being open, fast, and cost efficient.
Sebastian Scheurle, Regional Industry Lead Automotive EMEA at NVIDIA, said:
It’s been impressive and highly encouraging to see how post-trained Nemotron and SPREAD’s Engineering Harness technology outperforms leading frontier models on an important automotive engineering use case.
About SPREAD
SPREAD empowers global manufacturers to unite, understand, and act on their product data. Founded by Philipp Noll and Robert Göbel in 2019, SPREAD began with the belief that engineering decisions are only as good as the data behind them. That principle still guides how the company builds technology today. Leading global brands use SPREAD's Engineering Intelligence Platform to accelerate development, reduce costs, improve product quality, and increase uptime. SPREAD employs 100 people, primarily in their Berlin HQ. For more information, visit www.spread.ai.