Not tested by us yet. These are driving models for vehicles and simulators, not chat or API models, so our test harness cannot call them. Everything below comes from the TIER IV announcement on PR Newswire, and every figure is the vendor's. We'll update this post if independent driving benchmarks appear.
On 2026-11-11 TIER IV and the Autoware Foundation open-sourced five end-to-end driving models built to plug into Autoware 2.0. They span a 16 million parameter camera model (METEOR) to NVIDIA's 2 billion parameter Alpamayo 2 Edge, with different bets on maps, sensors and chips. The release publishes no benchmark scores and no licence, so the useful reading is a map of options, not a ranking.
| Model | Size | Inputs | Maps | Status per announcement |
|---|---|---|---|---|
| METEOR | 16M parameters | 1 to 8 cameras, optional LiDAR and radar | No HD maps | Released |
| AutoE2E | 79.9M parameters | Cameras up to LiDAR and radar setups | Not stated | Reactive Model released and integrated; World and Reasoning Models in development |
| Alpamayo 2 Edge | 2B parameters | Camera images only | Not stated | Autoware integration complete; optimisation for real vehicles ongoing |
| OnePlanner | Not stated | LiDAR (BEVFusion), HD maps, route, traffic signals, vehicle state | HD maps | Integration complete, 6 second trajectories |
| LatentDrive | Not stated | 1 or 3 cameras, route, vehicle state | No HD maps | Integration and in-vehicle validation underway, 4 second trajectories |
What each model is for
METEOR is a multitask CNN with a bird's-eye-view representation, so one network handles perception through trajectory generation. TIER IV aims it at privately owned passenger cars and says it runs on NVIDIA and Renesas chips. At 16 million parameters it is the lightest of the set, and the announcement says it is lighter than AutoE2E.
AutoE2E uses a fast-slow design: a Reactive Model for real-time decisions, a World Model that predicts how the scene changes, and a Reasoning Model that supports decisions. Only the reactive part ships today. TIER IV says it is considering robotaxi commercialisation and that the model runs on hardware beyond NVIDIA's.
Alpamayo 2 Edge comes from the NVIDIA collaboration. It is a vision-language-action model that reads camera images and reasons its way to trajectories. The announcement contrasts its 2 billion parameters with 34 billion for Alpamayo 2 Super, and the weights are on Hugging Face.
OnePlanner feeds perception features straight into the planner instead of converting them into object positions and velocities first. It uses HD maps and plans six seconds ahead.
LatentDrive encodes camera video into latent space with V-JEPA, a self-supervised method, and plans from there without object detection or tracking. It is the same world-model idea behind open video releases like PAN 2, pointed at a steering wheel.
What the announcement does not tell you
- No benchmarks. There are parameter counts and planning horizons, but no closed-loop driving scores, collision rates or latency figures.
- No licence stated. Check each repository before commercial reuse.
- Uneven maturity. Two models are fully integrated, one is partly released, and LatentDrive is still being validated in a vehicle.
- Vendor framing. TIER IV's pitch is that developers should not depend on one AI model or one chip vendor. That is a reasonable goal, but the claim has not been tested outside the company.
How it compares with other robotics releases
The closest comparison on paper is the vision-language-action line from NVIDIA itself, of which Alpamayo 2 Edge is a small member. For general robot control rather than cars, Google's Gemini Robotics 2.5 is a different kind of release, while these TIER IV models are weights you run in your own stack. Camera-only perception (Alpamayo 2 Edge, LatentDrive) trades sensor cost for harder perception, while METEOR and AutoE2E accept LiDAR and radar when you have them.
Who should care
- Autoware users and AV research teams. Five models behind one integration layer lets you swap approaches on the same vehicle or simulator, which is the real value here.
- Embedded teams on tight compute. METEOR's 16 million parameters and Renesas support are the numbers to check against your own board.
- Everyone else choosing models. This is a different category from the chat and decision models we usually test. If you are picking a model for text work, start with how to choose an LLM.
We'll retest nothing here until there is a simulator benchmark worth running. Retest condition: independent closed-loop results for any of the five.