FSD V14.3.2 Begins Rolling Out — One Unified AI Model for Smart Summon, FSD, and Robotaxi
4 min read read
Tesla started shipping FSD V14.3.2 on May 1, 2026. The initial rollout is small — roughly 2.7% of the eligible fleet, about 53 vehicles in the first batch — but the architectural change is the largest single jump since FSD V12 went end-to-end.
For the first time, the same AI model is driving Actually Smart Summon, supervised FSD on owner cars, and the unsupervised Robotaxi service in Austin.
What "Unified Model" Actually Means
Before V14.3.2, Tesla maintained three distinct neural network stacks:
- One model for Actually Smart Summon (low-speed, parking-lot navigation).
- One model for supervised FSD (consumer cars with a human ready to take over).
- One model for unsupervised Robotaxi (Austin-only, no driver).
Each was trained and tuned separately. Improvements in one didn't automatically transfer. V14.3.2 collapses all three into a single unified model. Training data from Robotaxi rides feeds the same model that handles Smart Summon. Edge cases caught in supervised FSD improve unsupervised driving immediately.
Specific Behavioral Improvements
| Area | What Changed |
|---|---|
| Traffic light handling | Compound lights, curved approaches, yellow-light decisions trained on hard reinforcement-learning examples from the fleet |
| Rare object response | Improved handling of objects extending, hanging, or leaning into the vehicle path (sourced from infrequent fleet events) |
| Vision encoder | Upgraded neural net for low-visibility scenarios, stronger 3D geometry understanding, expanded traffic-sign recognition |
| Reaction time | 20% faster, via a rewritten AI compiler and runtime built on MLIR |
The MLIR Rewrite Matters More Than the User-Facing Notes
Tesla rewrote the AI compiler and runtime from the ground up using MLIR (a modern compiler framework originally from Google's TensorFlow team). The headline benefit is the 20% reaction-time improvement. The second benefit is iteration speed — the engineering team can train, compile, and ship new model versions faster than before.
That's the multiplier. If model improvements ship in 4-week cycles instead of 8-week cycles, the V14.x line accumulates capability faster than the V12 line did.
"Honestly kind of freaky how humanlike." — Early V14.3.2 user feedback collected by Not a Tesla App
Who Gets It and When
The 2.7% rollout is normal for a major-ish FSD release. Tesla's typical staging:
- Week 1–2: Internal QA and a few hundred hand-picked beta cars.
- Week 3–4: 5–10% of the eligible fleet.
- Week 5–8: Fleet-wide if no regression metrics trigger.
If you're on V14.3.x already and haven't received .3.2, you're in the normal queue. If you're on V13.x or earlier, the upgrade path runs through 14.3.x first — expect it within the next two release cycles.
The Bottom Line
The headline ("FSD update") undersells what happened. Tesla just collapsed three separate AI systems into one. Every mile driven by a Robotaxi in Austin now improves the model that parks your car at the grocery store. That's a different kind of company than the one that shipped FSD V11 two years ago.
Photo: Tesla driving / Pexels