AUSTIN, Texas — Tesla released software update 2026.14.6.10, carrying FSD (Supervised) v14.3.4, on June 11, 2026. The update brings a ground-up rewrite of Tesla's AI compiler and runtime using MLIR — a change that reduces the FSD system's reaction time by 20% and accelerates future model iteration. It also delivers an upgraded neural network vision encoder capable of handling low-visibility scenarios more reliably, and completes the rollout of Actually Smart Summon (ASS) to Cybertruck owners. The update applies to all HW4 vehicles across the full Tesla lineup.
The MLIR Rewrite: Why It Matters
The most significant change in v14.3.4 may be the one least visible to drivers: a complete rewrite of the AI compiler and runtime from the ground up using MLIR (Multi-Level Intermediate Representation), an open-source compiler framework designed for machine learning workloads. Tesla reports that the MLIR implementation produces a 20% improvement in reaction time compared to the prior compiler stack.
In practical terms, faster reaction time means the FSD system processes its visual input and issues steering or braking corrections more quickly — a change that matters most at highway speeds or in fast-developing hazard scenarios. The MLIR transition also accelerates the pace at which Tesla can train and iterate future models, since better tooling compresses the engineering cycle between data collection and updated model deployment. As Tesla's cross-border European FSD deployment demonstrated this week, the underlying system is already active across five EU countries with an excellent safety record — a faster-reacting version of that system strengthens the safety data Tesla can bring to regulators in countries not yet approved.
Vision Encoder and Intersection Improvements
Alongside the compiler rewrite, v14.3.4 delivers an upgraded neural network vision encoder with three specific improvements: better handling of rare and low-visibility scenarios, stronger 3D geometry understanding, and expanded traffic sign recognition. Low-visibility is a well-documented edge case for autonomous systems — fog, glare, night rain, and backlit scenarios all challenge cameras in ways that clear-day highway driving does not. Targeting this explicitly with a revised encoder is a direct response to fleet-data-driven training that feeds harder examples back into the network.
