Nvidia released Alpamayo 2 Edge on 10 November, a 2-billion-parameter driving model that reasons about a scene and plans a path inside the car, its developer blog says. The weights are on Hugging Face under the OpenMDW-1.1 license, and the code under Apache 2.0.
In Nvidia’s closed-loop simulation test, the model drove 9.87 kilometres between at-fault incidents, against 0.25 for the 4-billion-parameter Qwen-Drive-1.0. Nvidia highlights that as 40x; the exact ratio is 39.5, by our arithmetic. The comparison is Nvidia’s own and we did not test the model.
Business Pill · API vs Open Weights
A short Business Engineer explainer on renting a model through an API versus running open weights yourself. It teaches the general idea only and does not discuss Nvidia or Alpamayo.
The key insight: Nvidia released Alpamayo 2 Edge, a 2-billion-parameter driving model with open weights under OpenMDW-1.1, built for its DRIVE AGX Thor computer. In Nvidia’s closed-loop simulation test it drove 9.87 kilometres between at-fault incidents against 0.25 for the 4-billion-parameter Qwen-Drive-1.0, a ratio of 39.5, by our arithmetic.
The model card calls it “a 2B-parameter reasoning vision-language-action model that uses adaptive reasoning to generate driving trajectories and causal explanations”. It is built on the backbone of Nvidia’s Cosmos 3 Edge and is meant for “efficient in-vehicle deployment on NVIDIA DRIVE AGX Thor”.
“Adaptive” means the model writes out its reasoning only some of the time. The blog says it reasons “generating reasoning only when needed”, and that it emits special tokens instead of a reasoning trace during routine driving. Nvidia’s post on X says it explains its decisions “when needed”.
The card lists a 2.45-billion-parameter backbone and a 0.2-billion-parameter action expert, 2.65 billion in total by our arithmetic. It takes six cameras, a text navigation instruction and the car’s recent motion, and returns 64 waypoints over 6.4 seconds with a written explanation.
Kilometres driven between at-fault incidents in Nvidia’s closed-loop simulation test: Alpamayo 2 Edge (2B parameters) against Qwen-Drive-1.0 (4B), as reported in Nvidia’s developer blog on 10 November 2026. Nvidia’s own benchmark; the 39.5x ratio is our arithmetic.
What Was Posted on X
NVIDIA Alpamayo 2 Edge is here, bringing adaptive AI reasoning in real-time to the vehicle. ๐
From yielding to pedestrians to making room for an oncoming bus, watch how it plans driving paths and explains its decisions when needed.
Developers can build on this open model to create the next generation of autonomous vehicles, with AI they can adapt to their own fleets.
Nvidia’s own account announced Alpamayo 2 Edge at 16:08 UTC on 10 November, saying the open model brings adaptive reasoning to the vehicle and explains its decisions when needed. The figures in this piece come from Nvidia’s developer blog and model card.
Speed and the Numbers Nvidia Gives
On DRIVE AGX Thor, the optimised model has “a pixel-to-trajectory inference latency of 160 ms to 190 ms depending on the amount of reasoning”, the blog says. The deployment recipe is public in Nvidia’s Alpamayo Developer Hub.
Post-training with reinforcement learning in Nvidia’s AlpaGym simulator improved closed-loop driving performance “by 48.9% on top of the best-performing checkpoint without closed-loop RL”. Training used roughly 115,000 hours of multi-camera driving video and roughly 3,700,000 reasoning traces, according to the card.
The card reports a separate closed-loop result of 15.15 kilometres per at-fault incident on 875 scenarios. It also says the model was tested on an Nvidia H100 GPU, with a peak of 12.06 GiB of GPU memory, and recommends at least 16 GB.
Who Has Used It
Nvidia’s blog quotes three early-access developers. Metas Simonavicius, CTO of PORT, said that in early simulation tests “Edge kept its lane in more than twice as many runs as Alpamayo 1.5”.
Luca Crotti of Niulinx said fine-tuning on the company’s own data “delivered a further 44.2% reduction in displacement error compared to the Alpamayo 2 Edge baseline”. Hans Salomonsson, CEO of Embedl, said: “In less than 48 hours, we made NVIDIA’s Alpamayo 2 Edge run at 10 Hz on AGX Thor for real-time planning.”
What Changed Since 21 October
On 21 October we reported Nvidia’s description of Alpamayo 2 Edge as a 2-billion-parameter model for DRIVE AGX Thor, alongside Bolt’s robotaxi plan. The model was announced then; now the weights, code and benchmark figures are published, 20 days later by our arithmetic.
Edge is the small sibling of the 34-billion-parameter Alpamayo 2 Super, which the product page describes as a model “for planning, auto-labeling, and evaluation”. The blog says Edge was trained with reasoning labels generated by Super. Our related coverage today: Nvidia’s GR00T 2 robot model and Google’s Gemini Robotics 2.5.
The Structural Read
As we read the blog, the design choice is about the onboard budget: the model writes out its reasoning only when needed, and Nvidia gives a 160 ms to 190 ms latency range depending on the amount of reasoning.
As we read the release, the weights are free but the deployment recipe Nvidia publishes is for its own DRIVE AGX Thor computer; the model card lists tests on an Nvidia H100 GPU.
Marco Pavone and Boris Ivanovic, Nvidia, developer blog, 10 November 2026
“Larger models provide a foundation for developing these capabilities, but moving them into a vehicle requires additional work to reduce model size, preserve useful reasoning, and validate performance on the target hardware.”
Three Implications
FOR AUTONOMOUS-DRIVING DEVELOPERS The weights are on Hugging Face under OpenMDW-1.1 and the code under Apache 2.0, the model card says.
FOR CARMAKERS AND ROBOTAXI FLEETS Nvidia’s X post says developers can adapt the model to their own fleets; Niulinx said fine-tuning on its own data cut displacement error by a further 44.2%.
FOR READERS OF THE BENCHMARKS The 40x figure is Nvidia’s own simulation comparison; the model card reports a separate 15.15 kilometres per at-fault incident on 875 scenarios.
The framework states the pattern as: “Close the SCARCE layer (keep proprietary) + Open the ABUNDANT layer (commoditize)”.
As we read it, Alpamayo 2 Edge fits that pattern: Nvidia opens the model weights and code, while the deployment recipe it publishes targets its own DRIVE AGX Thor computer. The framework predates this release and does not discuss it.
What Is Not Established
The benchmarks are Nvidia’s and its partners’, in simulation or early tests; none is a road-safety record. Nvidia does not explain the gap between the blog’s 9.87 and the card’s 15.15 kilometres, and names no carmaker putting Edge into production.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
Nvidia has published an open 2-billion-parameter driving model that runs on its own DRIVE AGX Thor computer and, in Nvidia’s simulation test, went 40x farther between at-fault incidents than a 4-billion-parameter rival. The weights and code are free to build on; the target chip is Nvidia’s.
A note on sourcing. Figures and quotations are from Nvidia’s developer blog of 10 November 2026, the Alpamayo 2 Edge model card on Hugging Face, Nvidia’s Alpamayo product page and Nvidia’s posts on X. Partner quotations are as published by Nvidia. The benchmarks are Nvidia’s; we did not test the model. Ratios, totals and day counts are our arithmetic.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
