PentaNet edges BitNet with pentanary quantization

BitNet's ternary alphabet is starting to look like a design point, not a ceiling; PentaNet suggests a slightly richer discrete space can recover capacity without abandoning add/shift-only inference. The caveat is that this is still a small, self-reported experiment, so the real proof will be scaling plus independent profiling.

• –The reported mean PPL drops from 192.63 to 180.32 across three seeds, which is a solid gain for a 124M language model.
• –The ±2 buckets staying occupied matters; if the distribution collapsed back to ternary, the extra states would just be dead complexity.
• –Add/shift-only arithmetic is the right constraint, but batch size, memory traffic, and compiler fusion will decide whether the speedup survives outside the demo path.
• –The repo already ships the training code, weights, and both Triton GPU and AVX2 CPU paths, so others can stress-test the claim.
• –If this holds up, it points to a broader design space of low-bit alphabets, not just ternary vs full precision.