Service eligibility and regional restrictions
PuppyIP serves only compliant overseas businesses and their authorized personnel. Proxy services are not available in mainland China. The service may only be used for lawful business activities outside mainland China. Use of this service within mainland China is prohibited.
Hosting a proxy IP or server overseas does not change these restrictions. The service must not be provided to end users in mainland China through relaying, forwarding, sharing or resale. Before use, read the Terms of Service.
Key Takeaways
- Tencent AI’s official account announced the open-source release at 21:55 Beijing time on September 4, 2026, and said the model already handles real WeChat traffic. These are Tencent’s disclosures, not third-party production validation.
- The official repository offers 2B, 4B and 9B models supporting text, images, video, visual documents and interleaved multimodal inputs. Audio is currently unsupported.
- In the official MMEB-v2 evaluation, the 2B model at 256 dimensions retained 98.7% of its full-dimensional image and video performance. Business selection still requires retesting with your own queries and documents.
- Pin the Git commit, Hugging Face revision and file hashes before downloading. Review remote model code first and run it in an isolated environment.
- A fixed egress can make repository and model downloads more reproducible, but it cannot fix insufficient VRAM, incompatible dependencies, preprocessing differences or retrieval-quality problems.
- Retain the old vector index and model before launch. Start with shadow evaluation, then switch a small share of traffic instead of immediately overwriting a large index irreversibly.
Official scope: announced at 21:55, with performance figures treated as publisher results
Tencent AI’s official account posted at 21:55 Beijing time on September 4, 2026 that WeMM-Embedding processes 1 billion requests daily in WeChat and had been open-sourced. The post also reported a score of 80.6 for 9B on MMEB-v2 and 98.7% performance retention for 2B at 256 dimensions. The official repository now provides model weights, inference examples, serving examples and evaluation code.
These figures come from Tencent’s post, repository and paper. They cannot be directly converted into your accuracy, throughput, cost or service level. MMEB-v2 is not your business dataset, and WeChat’s live traffic does not establish an open-source deployment configuration. Label the claims as “Tencent reports” when planning the project, and base adoption on local benchmarks and staged-rollout results.
Choose the input scope before choosing 2B, 4B or 9B
The official repository lists 2B, 4B and 9B models with full vector dimensions of 2048, 2560 and 4096 respectively. All three support text, images, video, visual documents and sequentially interleaved multimodal inputs. Audio is currently unsupported. First determine whether the task is text-to-image retrieval, image-to-product retrieval, video clips, visual documents or mixed-content retrieval. Do not connect every data source simply because the model is called general-purpose.
Run a full-dimensional 2B baseline on the same de-identified samples before deciding whether 4B or 9B is needed. Record model weights, input sizes and frame sampling, batch size, peak VRAM, single-item and batch latency, index size, retrieval metrics and failure rate. Without comparable results, parameter counts or public leaderboards alone cannot support a model-selection decision.
Trusted downloads: pin revisions and hashes, and review code
Obtain code and weights only from Tencent’s official GitHub organization and the tencent organization on Hugging Face. Save the Git commit, model revision, download time, file inventory and SHA-256 hashes, and use the model cache as a read-only build input. Do not automatically fetch a moving main branch or the latest model files at every startup. The repository says Tencent code uses Apache License 2.0 by default, while third-party components retain their own licenses. Check the model card, LICENSE and dependency licenses before commercial use.
Model loading may execute Python code included in the repository. Review and run a pinned revision in an isolated environment with no production keys, no customer data and restricted network access before creating an internal image or artifact. Store Hugging Face tokens, proxy credentials and object-storage credentials only in secret management, never in scripts, image layers, logs or tickets.
Troubleshoot download networks separately from the inference environment
When repository or weight downloads fail, record the target domain, file, revision, downloaded bytes, HTTP status, DNS, TLS, timeouts and proxy authentication results. Use resumable downloads and a local cache, and check temporary files and final hashes before retrying. For stable cross-region access to GitHub or Hugging Face, use the proxy connection checklist to validate each layer. Fixed egress supports consistent access and reproducibility; it should not be used to bypass platform terms, model licenses or regional restrictions.
If the model has downloaded completely but fails to load, check Python, PyTorch, CUDA, transformers, VRAM and preprocessing code. For reproducibility, the official repository recommends transformers==5.2.0 and notes that newer versions may produce preprocessing differences. Insufficient VRAM, incompatible operators, shape errors and abnormal retrieval scores are not network issues that changing a proxy can fix.
Evaluate Matryoshka dimensions with retrieval, latency and index costs together
The official implementation supports taking the leading components of a full vector at listed Matryoshka dimensions, followed by L2 renormalization. The 2B model supports 64, 128, 256, 512, 1024 and 2048 dimensions; 4B supports up to 2560; 9B also supports 4096. Do not truncate to arbitrary unlisted lengths or omit renormalization, as similarity distributions and thresholds will no longer be comparable.
Build a two-stage experiment: use 64 or 256 dimensions for fast retrieval of a large candidate set, then rerank those same candidates with 1024, 2048 or full dimensions. Compare this with single-stage full-dimensional retrieval using Recall@K, NDCG, false matches, index memory, query latency and rebuild time. The official 98.7% figure applies to specific MMEB-v2 image and video results and does not guarantee the same retention for your products, ads, documents or Chinese-language queries.
Validate input order, frame sampling and interface contracts before serving
Prepare positive examples, hard negatives and missing inputs separately for text, images, video and visual documents. Then add interleaved multimodal samples and confirm that changes in input order match business expectations. Before indexing vectors, save the model version, dimension count, normalization method, preprocessing version and content version. Do not mix vectors from different models or dimensions in the same unpartitioned index.
The official repository provides examples for Transformers, Sentence Transformers, vLLM 0.27.0 and SGLang 0.5.9, but official testing does not establish compatibility with your hardware and gateway. At minimum, the server needs readiness checks, timeouts, batch limits, input-size and video-frame limits, queue length, VRAM levels and version information in responses. Keep de-identified request IDs for failures, and avoid letting automatic retries amplify inference-queue congestion.
Staged rollout and rollback: retain the old model, index and query logs
First replay a batch of de-identified queries offline, generating results from the old model and WeMM-Embedding in parallel without affecting users. After manual labeling and business metrics confirm a benefit, send a small proportion of read-only search traffic to the new index. Recommendation, search and ecommerce flows should also check popular-content bias, language and region distribution, duplicate content, cold starts, tail latency and the no-results rate.
Keep the old model artifact, old index snapshot, versioned routing and a reversible switch during cutover. Stop expanding the rollout immediately if hashes differ, remote code has not been reviewed, dimensions are mixed, retrieval declines, latency or VRAM exceeds limits, queues keep growing, input order is wrong or logs contain sensitive content. Switch back to the old route first, then isolate the problematic version. Do not overwrite the entire vector store in place during an incident.
Sources
Frequently Asked Questions
Which inputs does WeMM-Embedding support?
The official documentation lists text, images, video, visual documents and interleaved multimodal inputs. Audio is currently unsupported. Validate specific formats, sizes and video sampling against the pinned model card and code.
Should we choose 9B immediately?
Do not decide only from a public leaderboard. First record retrieval quality, latency, VRAM and index costs for 2B on your own de-identified data. Upgrade only when the gains from 4B or 9B exceed the resource and operational costs.
Will 256 dimensions always retain 98.7% of performance?
No. The 98.7% figure is Tencent’s reported 2B result for MMEB-v2 image and video tasks. Your data, query distribution, metrics and thresholds differ. Retest locally and renormalize after truncation.
Can remote model code be used directly in production?
Directly loading a moving version is not recommended. Pin the revision, review the code and test in isolation before creating reproducible internal artifacts and minimizing network and secret permissions.
Should we change proxies first when a download fails?
First record HTTP, DNS, TLS, timeouts, proxy authentication, revision and file hashes. Adjust egress only when evidence points to the network layer. Dependency, CUDA, VRAM, preprocessing and vector-quality issues belong in the inference environment.
Why keep the old index during launch?
Vectors produced by different models, dimensions and preprocessing cannot be mixed freely. Retaining the old model, index and routing switch allows a quick rollback when retrieval declines, dimensions are mixed or performance exceeds limits.