Aaron

What's the best concert you've ever been to?

Following future rollout of FSD V14 Lite for HW3 vehicles in the US, we plan on expanding V14 Lite to additional international markets. This update ensures that HW3 vehicle owners will continue to benefit from ongoing software updates. Since international rollout is subject to several factors (completion of technical verification, regional adaptation & relevant regulatory approvals), we can't provide definitive dates at the moment, but will provide updates on a rolling basis

Tesla could use several proven techniques to shrink a large FSD model for older HW3 (and even HW2.5) while keeping latency close to HW4: - **Quantization**: Drop from FP32/FP16 to INT8 or INT4 weights/activations. Tesla already does heavy int8 inference. - **Pruning + Sparsity**: Cut 30-50% of weights that contribute little, then use sparse kernels optimized for their NPU. - **Knowledge Distillation**: Train a smaller "student" net to mimic the big "teacher" model's outputs. - **Layer Fusion / Operator Fusion**: Merge ops (conv+BN+ReLU) and rewrite for the exact HW3 silicon. - **Dynamic Early Exit**: Let easy frames exit early in the network. They've done similar for v12.x already. v14 lite is likely this + custom compiler tweaks. HW3 should handle it fine with those. 🫡