paper web signal

WeChat WeMM-Embedding 2B beats prior 8B on MMEB-v2, 9B hits 80.6

TL;DR

  • The 9B WeMM-Embedding variant reports a new MMEB-v2 state-of-the-art overall score of 80.6.
  • The 2B variant claims to already surpass the previously leading 8B open-source baseline on MMEB-v2.
  • WeChat says the models are backed by a 26-task in-house benchmark and 14 online A/B tests.

WeChat's engineering team says the 9B variant of its new WeMM-Embedding family hits a new state-of-the-art overall score of 80.6 on MMEB-v2, according to a technical report on arXiv. The smaller 2B variant, the report claims, "already surpasses the previously leading 8B open-source baseline on MMEB-v2."

The family covers text, images, videos, visual documents, and arbitrarily interleaved multimodal inputs, in 2B, 4B, and 9B sizes with flexible output dimensions. Training runs in two stages: a large-scale multimodal alignment pass, then a refinement stage using curated data, fine-grained relevance supervision, and cross-scale knowledge transfer.

Alongside the public numbers, the authors point to internal evidence, reporting "substantial gains on a 26-task in-house benchmark and consistent improvements across 14 online A/B tests" across WeChat properties.

The abstract does not publish per-variant MMEB-v2 scores, per-task deltas, the identities of the in-house tasks, or the name of the 8B baseline that the 2B is said to beat.