LongCat 2.5 W4A4 released

LongCat 2.5 W4A4 is a compressed copy of Meituan's LongCat 2.5 model in which the expert layers compute with 4-bit integers instead of 16-bit numbers. The measurements Meituan publishes for it are serving-speed gains over an 8-bit version, not over the full-precision model. Its model card is the same text as the card for the full LongCat 2.5 release, so it does not report separately how this variant scores on benchmarks.

A model's weights are the numbers it learned in training, and its activations are the intermediate values it calculates while processing each token. The full release stores its weights in BF16, a 16-bit format. W4A4 means both the weights (W) and the activations (A) are rounded to 4-bit integers, a format called INT4 that can hold only 16 distinct values. Rounding that coarsely usually costs accuracy, so Meituan says it used learnable transformations in the style of FlatQuant, a published method that reshapes each layer's values before rounding so a few extreme values do not distort the 4-bit scale. Because the activations are 4-bit as well, the multiplications themselves can run as INT4 operations on hardware that supports them, which saves computation as well as storage. The repository's configuration file applies the 4-bit setting to the expert layers of the model's mixture-of-experts design, uses 8-bit integers for some other layers and leaves parts such as the input embedding and output layers unquantized. Hugging Face's file listing puts the weight files at about 1.3 terabytes, against about 3.6 terabytes for the BF16 release.

The speed figures compare this version with a W8A8 version of the model, in which weights and activations use 8 bits. In what the card calls online stress tests, Meituan reports 43.3% higher prefill throughput (prefill is the stage where the model reads the prompt), 42.77% higher decode throughput (decode is generating the reply one token at a time), 15.91% less time before the first token appears and 19.72% less time per output token. Adding an 8-bit KV cache, the stored record of earlier tokens that the model consults while generating, brought decode throughput to 92% above the 8-bit setup with a 16-bit cache. These are Meituan's own numbers. The card does not name the hardware, request load or prompt lengths behind them. It says only that the 4-bit path runs natively on "domestic AI accelerators", and its deployment section covers only NPU platforms, a general term for chips built for AI workloads.

On quality, the card says only that the 4-bit model keeps accuracy comparable to the unquantized model on the benchmarks Meituan evaluated. It does not name those benchmarks or give the scores, and its large benchmark table does not say whether it was produced with this version or the BF16 one. What the release establishes is a much smaller download and a set of vendor-measured speedups against an 8-bit baseline. Whether the 4-bit model's output matches the full model's on a given task is a claim that still needs independent checking.