A new 4.5-bit number format kept three vision-language models close to a leading 4-bit baseline while its synthesized compute array delivered 25% more computation per square millimeter. The result comes from a MICRO 2026 paper on MiX, a format that changes how the intermediate numbers produced by the model are represented so the accelerator can replace most multipliers with shifters. The hardware figures come from a simulated 28-nanometer design and estimates from its hardware description, not a fabricated chip.
Vision-language models must compress intermediate numbers from images and text that can differ greatly in size. A common low-bit format makes a block of numbers share an exponent, the part that sets their scale. One large value can force the block onto a scale too coarse to preserve its smaller values. Compressing only the model's fixed weights avoids that problem, but leaves much of the working data expensive to store and move.
MiX instead shares a mantissa, the significant digits, while giving each number its own small scale adjustment. The shared part can be multiplied once after summing a group's contributions. Much of the remaining multiplication becomes shifts of binary digits and integer addition. The format thus changes both which numerical detail is retained and which circuits the accelerator needs.
Attention, the operation that selects relevant information from the input, needs an additional correction. MiX removes a common offset from its stored key values before compressing them, making extreme values less damaging. It calculates that offset while processing the prompt and reuses it as the answer is generated.
Across six visual reasoning, reading and question-answering benchmarks, the 4.5-bit MiX configuration scored slightly higher than NVFP4 on LLaVA-OneVision and MiniCPM-V, and slightly lower on Qwen2-VL. These were averages across the six tests, not a single task's success rate.
For the compute array alone, synthesis at 500 megahertz reported 9.1 tera-operations per second per square millimeter and 16.1 tera-operations per watt. The comparable NVFP4 array reached 7.3 and 14.0, respectively. A cycle-level full-system model then showed MiX running 2.3 to 4.5 times faster and using less energy than the Focus token-pruning accelerator across the tested workloads. Those system results remain modeled comparisons: latency and active energy were derived from the hardware-description code, memory behavior came from simulation, and long decoding sequences still approached the same memory-bandwidth ceiling as other low-bit formats.
MiX therefore offers evidence that the number format and datapath can be designed together to make end-to-end low-bit vision-language inference more practical. The authors released an MIT-licensed artifact with code, hardware descriptions and synthesis scripts, but the paper does not report silicon measurements, and reproducing the hardware numbers requires proprietary synthesis tools and a process library. Text-only language-model tests were also less favorable than the vision-language results.
