Hero title card reading 'WeMM-Video-2B', subtitled 'shipped without an announcement', with four badges reading '2B dense, 3.28B total', 'video + audio + image + text in', 'Apache 2.0, four quantisations', 'zero downloads on the Hub'.
Engineering & Research

WeMM-Video-2B: A 2B Omni-Modal Model Nobody Announced

Author

Gideon Frost

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

There is no launch post. There is no model card on a product page, no waitlist, no pricing table, no benchmark tweet from a T​encent executive. What exists is a Hugging Face repository created at 08:58 UTC on 17 November 2026 and a GitHub repo whose first commit landed six and a half hours earlier, and both of them describe WeMM-Video-2B as the dense 2B checkpoint in a family of Chinese–English omni-modal video models from WeChat Vision at T​encent. Six days later the repo has two likes, zero downloads, and four quantised siblings that were created within eleven minutes of the base weights. So the honest first sentence of any piece about this model is a sentence about its packaging: it is public, it is downloadable, and nobody has said a word about it.

That gap is worth taking seriously rather than treating as a slow news day. A model family that ships weights, a paper, four quantisation formats, patched vLLM and SGLang recipes, a native C++ runtime and a Docker compose file in the same week is not a research dump someone forgot to announce — it is a release that was prepared and then released without the marketing layer. What follows separates what the artifacts prove from what only T​encent can confirm, and it is deliberately uninterested in deciding whether the model is good.

What the repository actually contains

The base checkpoint is tencent/WeMM-Video-2B, Apache 2.0, 6.12 GiB of BF16 safetensors across two shards. The index file reports 3,283,624,264 parameters in total, which tells you something the name does not: "2B" is the language model, not the system. The text backbone is a Qwen3.5-family decoder — 24 layers, hidden size 2048, 8 attention heads against 2 KV heads, a 248,320-token vocabulary and a 262,144-token maximum position embedding — with a hybrid attention layout that puts one full-attention layer after every three linear-attention layers. Sitting in front of it are a 24-layer vision transformer at hidden size 1024, and an audio stack built from a 32-layer Whisper encoder feeding a two-layer Q-Former with a hidden size of 4104. The remaining parameters live in those two encoders.

Generated scoreboard titled 'WeMM-Video-2B — the scoreboard' with six rows: parameters 3.28B total with a 2B language model, modalities in text, image, audio and video plus soundtrack, output text only, video sampling 2 fps from 4 to 16 frames, an 8,192-token context, and no hosted endpoint as the price. A footer reads 'Vendor-reported; no independent evaluations published.'

That architecture answers the question the name raises. This is not a video generator. It takes video in — frames plus, optionally, the original soundtrack — and produces text out. The config's own token IDs confirm it: a video token, an audio token, an image token, and a language model that generates. If you arrived hoping for another entry in the text-to-video race, the repository will disappoint you on the first read of the pipeline tag, which is image-text-to-text.

• Sizes in the family — 2B, 4B, 9B dense, and a 35B-A3B mixture-of-experts, all published together on 17 November.

• Quantisations for the 2B — FP8 Block128, AWQ W4A16, GGUF Q4_K_M and NVFP4 W4A16, each in its own repository, each created within eleven minutes of the base weights.

• Serving paths — Transformers 5.2.0, patched vLLM 0.18.0 or 0.31.0, patched SGLang 0.5.16 or 0.5.19, and a native runtime for the GGUF build that runs on CPU or CUDA.

• Input limits in the shipped recipes — up to four images, two videos and two audio items per request, at an 8,192-token context length.

• Licence — Apache 2.0 on the weights, with a separate third-party notices file covering LAVIS, vLLM and SGLang components.

The patched-serving detail is the one that separates this from a typical open-weights drop. T​encent did not ship a "works in theory" repository; it shipped overlay patches for specific vLLM and SGLang versions, locked dependency files, Docker images tagged to CUDA 12.9, and a validation note saying the checkpoint loaded and completed generation on L20 and H20 hardware across text, image, audio and video-with-soundtrack. Whether those patches hold up outside T​encent's own cluster is exactly the kind of question that a week of strangers running them answers, and that week has not happened yet.

Screenshot of the Hugging Face model card for tencent/WeMM-Video-2B, showing the image-text-to-text pipeline tag, Transformers and Safetensors, the qwen3_5 architecture with custom code, an Apache-2.0 licence, a model size of 3B parameters, a downloads figure with no deployments, and a Quantizations section listing the four sibling repositories.

What the paper claims, and who checked it

The technical report is arXiv 2611.26409, dated 19 November 2026 — two days after the weights appeared — and it is a family paper, not a 2B paper. Twelve authors, first-listed Changli Tang, last-listed Jing Lyu, all under the WeChat Vision banner.

The headline claim is about the 35B-A3B model, not the 2B: that it establishes a new open-source state of the art in omni-modal video evaluation, substantially outperforming Qwen3.5-Omni-Flash and Seed2.0-Mini-0428 and narrowing the gap with Gemini-3.0-Flash. A second claim does concern the model in this article's title: that WeMM-Video-9B surpasses the larger open-source Qwen3-Omni-30B-A3B-Instruct. The 2B appears in the family description and in the model zoo, but the paper's argument is carried by its biggest sibling.

Every one of those numbers is vendor-reported. There is no Artificial Analysis entry, no LMArena result, no independent reproduction, and — as of this writing — no third-party evaluation of any kind for any member of the family. The paper says the models were evaluated across more than 120 benchmarks; it does not say who ran them, and the answer is T​encent. That is not an accusation, it is the normal state of a six-day-old release. But it does mean the correct description of the 35B-A3B's position is "T​encent reports that it leads", not "it leads".

One further caveat is structural rather than editorial. A benchmark suite of 120 tasks, chosen and weighted by the lab that trained the model, is a measurement of fit — it tells you the model is good at the things its authors thought to measure. It does not tell you it is good at the things you need. The gap between those two statements is where most model-selection mistakes are made, and no amount of table-reading closes it.

The four modes, and why the audio path matters

The inference API exposes five request shapes, and the interesting one is the last. Text and image are unremarkable. Audio alone is unremarkable. Video frames alone is the standard multimodal-video case. Video with its original soundtrack is the one that changes what the model can be asked to do, and it is a separate content type in the serving API rather than an implicit behaviour.

In practice that means you can hand the model a clip and ask it to describe what is on screen and summarise what is being said, in one call, and get an answer that reflects both. The model card's own examples do exactly that — transcribing English speech from an Apollo 11 recording, then describing the same MP4 with and without its audio track. Whether the joint reasoning is better than running a video model and a speech model side by side and stitching the outputs is an empirical question, and it is the single most useful thing an independent evaluator could test.

The other design decision worth naming is the two-mode chat template. Passing enable_thinking=False gives a direct response; enable_thinking=True gives a reasoning trace before the answer. That is a latency-versus-quality dial shipped as a template flag rather than as two checkpoints, which is a reasonable choice for a 2B model — you are not paying for a second set of weights to get the second behaviour.

Running it, and the honest state of availability

The quick-start path is a Python 3.12 environment, an NVIDIA GPU, a driver compatible with CUDA 12.9 and FFmpeg. Install Torch 2.10.0, TorchVision 0.25.0 and TorchCodec 0.10.0 from the CUDA 12.9 index, then Transformers 5.2.0, Accelerate, Pillow, Librosa, SoundFile, SciPy and Requests. The model loads through AutoModelForImageTextToText with trust_remote_code=True — the architecture is custom, so that flag is load-bearing rather than decorative, and the card correctly tells you to read the custom code before enabling it.

If you want a server rather than a script, the vLLM recipe pins a patched 0.18.0 build and passes --no-enable-prefix-caching alongside --mamba-ssm-cache-dtype float32, which is a combination that tells you the hybrid linear-attention layers do not yet have a working prefix cache. That is a real throughput limitation for chat workloads, and it is the kind of detail that only surfaces when someone has actually tried to serve the thing.

What is absent from every document is a price. There is no Tencent-hosted endpoint for WeMM-Video-2B, no per-token rate, no managed deployment. You download the weights and you own the GPU. On the OrcaRouter side the honest statement is the same one: we do not serve WeMM-Video-2B. What we do carry from T​encent is the Hunyuan text line — tencent/hy4, tencent/hy4-flash and the preview builds — on one OpenAI-compatible endpoint at provider list price with 0% markup, which is useful if your pipeline needs a text model from the same lab and useless if it needs this one. Two different products, one vendor name, and conflating them would be the easy mistake to make in a hurry.

Screenshot of the OrcaRouter catalogue page for tencent/hy4, showing pricing of $0.56 per million input tokens and $2.24 per million output tokens, a quality score of 90, a speed score of 67, a 1M-token context window, a 64k maximum output, zero-data-retention support and a release date of 4 November 2026.

From curiosity to recommendation

Three things, in order of how much they would move the needle.

The first is an independent evaluation on a benchmark T​encent did not choose. Any blind arena run — video understanding, audio-visual QA, whatever is closest — would immediately tell you whether the family's claimed position survives contact with a neutral prompt set. Until that exists, the 35B-A3B's "new open-source state of the art" is a claim about a table in a PDF.

The second is a serving report from outside T​encent. The patches are specific, the hardware is specific, and the validation note covers two GPU models. Whether this runs cleanly on an H100 node, an A100 node, or a rented 4090 is unknown, and the answer will determine whether the 2B is a weekend project or a research artifact.

The third is a vendor announcement — any vendor announcement. A blog post, a model card on a T​encent product page, a pricing tier. Right now the release is legible only to people who read repository timestamps, and the seven-day window in which that is true is closing. If T​encent publishes nothing, the most likely reading is that these weights exist to support the paper and the 35B, and the 2B is a convenience checkpoint rather than a product.

Which is the honest bottom of it. WeMM-Video-2B is a real, downloadable, Apache-2.0 omni-modal model that nobody has announced, nobody has downloaded, and nobody has independently tested. If your work involves joint audio-visual understanding and you have a spare GPU and an afternoon, the cost of finding out what it does is the download. If your work involves shipping something next month, the cost of finding out is higher than that, and the responsible thing is to wait for someone else's numbers.

Compared in this article1

Detected from this article