Glonce

Tencent's WeMM-Embedding beats 8B rivals with a 2B model

modelsThe Glonce editorsHugging Face🔥55Glonce rating55Impact65Novelty60Relevance100Surprise35Players70Credibility88
Tencent's WeMM-Embedding beats 8B rivals with a 2B model

Tencent has published a technical report for WeMM-Embedding, a family of universal multimodal embedding models built to represent text, images, videos, visual documents and arbitrarily interleaved multimodal inputs in a shared vector space, with flexible output dimensions. The family comes in three sizes: 2B, 4B and 9B parameters. Training runs in two stages: a large-scale multimodal alignment stage first, followed by a refinement stage that uses curated data, fine-grained relevance supervision and cross-scale knowledge transfer between the different model sizes.

On public benchmarks, the 2B variant already surpasses the previously leading 8B open-source baseline on MMEB-v2, meaning a much smaller model outperforms a rival roughly four times its size. The 9B variant goes further, reaching a new state-of-the-art overall score of 80.6.

The models were also tested inside Tencent's own products rather than only on academic benchmarks. On a 26-task in-house benchmark, WeMM-Embedding shows substantial gains, and it produced consistent improvements across 14 separate online A/B tests. It is now deployed at scale in recommendation and search across WeChat Channels, WeChat Official Accounts, Moments, and e-commerce services.

Tencent has released the model weights and code on GitHub, under the Tencent organization, to support further research.

Key facts

  • WeMM-Embedding is a family of multimodal embedding models in 2B, 4B and 9B sizes, supporting text, images, video, visual documents and interleaved multimodal inputs.
  • Training uses two stages: a large-scale multimodal alignment stage, then a refinement stage with curated data, fine-grained relevance supervision and cross-scale knowledge transfer.
  • The 2B variant surpasses the previously leading 8B open-source baseline on the MMEB-v2 benchmark; the 9B variant sets a new state-of-the-art overall score of 80.6.
  • In production testing, the models show substantial gains on a 26-task in-house benchmark and consistent improvements across 14 online A/B tests.
  • WeMM-Embedding is deployed at scale in WeChat Channels, WeChat Official Accounts, Moments and e-commerce search and recommendation; weights and code are released on GitHub.

Why it matters

Multimodal embeddings are becoming core infrastructure for retrieval, recommendation, classification and agentic systems, since they let text, images, video and documents be compared in one shared space. The headline result here is efficiency: a 2B-parameter model beating the previous best 8B open-source baseline means comparable or better quality at a fraction of the compute and serving cost, which matters for anyone running embeddings at scale.

Who it affects

Teams building search, recommendation or retrieval systems that need to handle mixed content types (text alongside images, video or documents) are the direct audience, along with researchers working on multimodal representation learning. The WeChat deployment also makes it relevant to anyone studying how large platforms validate models in production, not just on academic leaderboards.

How to use it

Tencent has released the model weights and code on GitHub under its own organization account, so the models can be downloaded and run directly. The source does not give a release date, version number or licensing terms beyond pointing to the GitHub repository, and it gives no price or paid-tier information, so nothing beyond a public open release can be stated with confidence.

How solid is it

The claims come from Tencent's own technical report, evaluated on public benchmarks including MMEB-v2. The report adds two forms of internal validation beyond the academic score: a 26-task in-house benchmark and 14 separate online A/B tests inside live WeChat products, which is a more thorough production check than most model releases publish. The report does not name individual authors or the specific research group behind it beyond the Tencent GitHub organization, and it gives no benchmark figure for the 4B variant specifically.

Risks and caveats

The performance claims are self-reported by the team that built the models, without independent third-party benchmarking cited in the source. The comparison for the 2B variant is against a single class of baseline, the previously leading 8B open-source model, rather than the full field of competing embedding models. The report also omits architectural details such as the model backbone and the size of the training data, describing the refinement stage only as using "curated data" and "fine-grained relevance supervision" without further specifics.