What the Numbers Show—and What They Do Not

BitNet b1.58 2B4T is the point at which the native low-bit design becomes a measurable system claim. This chapter reads the reported results across task capability, memory, latency, and energy, then compares the model with INT4 and other near-1-bit alternatives. The useful judgment is not whether one column is highest; it is whether the model delivers a credible capability-per-resource trade-off under the stated benchmark, hardware, and evaluation conditions.

The main 2B4T comparison

The 2025 technical report compares BitNet b1.58 2B4T with instruction-tuned open-weight models between roughly 1 billion and 2 billion parameters. The table puts models from different families on a common evaluation run and reports both ordinary capability measures and operational metrics. That combination matters. If the table contained only benchmark scores, it would say little about the deployment problem that motivated BitNet. If it contained only memory and latency, it would not establish that the smaller numerical path still produces a useful language model.

The resource figures are the clearest part of the result. BitNet b1.58 2B4T is listed with 0.4 GB of non-embedding memory, 29 milliseconds of CPU time per output token, and estimated energy of 0.028 joules. The comparison models use between 1.4 and 4.8 GB of non-embedding memory, show 41 to 124 milliseconds per token, and have estimated energy values from 0.186 to 0.649 joules. On the face of those measurements, the native low-bit model occupies a much smaller memory footprint and decodes faster on the reported CPU path.

The qualifier “non-embedding” is essential. It means the memory number excludes the embedding component, so it should not be read as the amount of every byte in a complete deployment. The representation also has to coexist with activations, scales, runtime buffers, and the implementation’s packing choices. The figure is still useful: it compares the part of the model where the low-bit weight design is expected to have its largest effect. But it does not turn the nominal 1.58-bit label into a universal memory multiplier.

The latency number has an equally specific meaning. The authors measured average time per generated token after producing 128 tokens on a Surface Laptop Studio 2 with a 13th-generation Intel Core i7-13800H and eight CPU threads. BitNet used bitnet.cpp, while the other models used llama.cpp. That is a direct test of the released inference paths under a defined setup. It is not a promise that every CPU, GPU, batch size, sequence length, or serving framework will reproduce 29 milliseconds. Latency belongs to the model-plus-kernel-plus-hardware combination.

The energy figure has a different status again. The report estimates arithmetic-operations energy using an energy model associated with a 7 nm process node and a sequence length of 512. It is therefore an estimate of a defined part of the computation, not a direct battery or wall-power measurement for an entire device.