BitVLA (arXiv:2506.07530, code released as ustcwhy/BitVLA) is the first fully native 1-bit vision-language-action model -- every parameter in {-1,0,+1}, no post-hoc rounding. It inherits BitNet b1.58 2B4T as its language backbone and adds a Quantize-then-Distill stage that compresses the vision encoder to 1.58-bit weights with INT8 activations after multimodal training -- quantization-AWARE training, not post-training quantization, exactly module 2's recipe (a non-learnable absmean scale, train through the snap) scaled from a 96-weight XOR net to a real robot policy. The reported numbers, EXTERNAL and drawn directly from the paper: a 1.4 GB footprint, 11.0x smaller than the full-precision OpenVLA-OFT baseline; 4.4x faster end-to-end latency; and task performance that MATCHES that full-precision baseline on the LIBERO benchmark suite and real-world manipulation tasks -- not 'close to,' matches. This is lesson 5's claim, confirmed by someone else, on a policy in the Institute's own domain.
Now apply the arithmetic discipline this course has trained into you. It is tempting to multiply 11.0x by 4.4x and announce a '48x total win.' Don't. Memory and latency are different resources measured in different units -- bytes and seconds -- drawn from different bottlenecks (a smaller weight footprint vs. a faster decode-and-inference pipeline) that partially share a cause but were never re-derived as one joint pipeline. BitVLA's own paper never composites them into a single figure, for the same reason 'the honest joules' would not let a deleted multiply and a 20x-smaller weight compound into a mythical 1000x. Report the two ratios separately, cited, or not at all.
Now place our own number next to BitVLA's, honestly. Lesson 7 measured a ternary decode kernel running 5.8x faster and 3.2x more energy-efficient than bf16 on an RTX 4050. That number is INTERNAL -- Charlot Lab, unreplicated outside the Institute -- and it is MEASURED ON COMPUTE, not actuators: a decode microbenchmark on a desktop GPU, not a trained policy driving a robot arm. BitVLA's 11.0x/4.4x is EXTERNAL, published with released code and checkpoints, and measured end-to-end on a real manipulation policy. These two numbers are allies, not equals, and neither belongs beside the other without that flag.
Here is the frontier framing BitVLA earns us the right to make. The Institute's ownable slot in the ternary/energy-based landscape is a five-way intersection, and BitVLA closes exactly ONE leg of it: native ternary weights, in production, at scale, with parity proven. It does not claim, and does not need to claim, the other two legs this course's sister research stakes out -- an energy objective where the SAME scalar minimized to pick an action doubles as a formal stability certificate, and a worst-case contraction/bound-propagation GATE checked before a robot commits to a move, not an average-case benchmark score like a LIBERO success rate. Bringing all three together is still uncommon in the published work so far: ternary weights AND an energy-as-certificate objective AND a pre-commit worst-case gate, continuously learned. That is not a knock on BitVLA; it is the honest map of an open frontier.