A title card headed 'WeMM-Video-4B' with the subtitle "Tencent's audio-visual model, on Hugging Face without a launch post", three cards reading 'Name on the card 4B / 5.61B actual parameters', 'Repositories 18 / created in 19 minutes' and 'Hosted price none / self-host only', and a footer reading 'Figures from the Hugging Face repo and arXiv 2611.26409; no independent benchmarks.'
Engineering & Research

WeMM-Video-4B: Tencent's Quiet 5.6B Audio-Visual Model

Author

jinhao.song

Date Published

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

The file labelled WeMM-Video-4B weighs 10.45 GiB and, at two bytes per BF16 parameter, that works out to 5.61 billion parameters — not four. T​encent's own safetensors metadata says the same thing: 5,611,746,312 parameters. The "4B" in the name is the language backbone, not the checkpoint. That small arithmetic slip is the right way in to the WeMM-Video family, because almost everything interesting about this release is a gap between the label and the thing: a four-size family published with no launch post, a benchmark suite that beats Qwen3-Omni-30B-A3B-Instruct and MiniCPM-o-4.5-9B on video while losing to Qwen3.5-Omni-Flash on audio, and a 256K position embedding that every shipping configuration voluntarily cuts to 8,192 tokens. The weights are real, the paper is public, and nobody has downloaded any of it yet.

What follows separates three things that the family's Hugging Face cards run together: what is measurable in the repository, what only T​encent's own evaluation reports, and what has not been confirmed by anyone outside WeChat Vision.

The file called 4B holds 5.61 billion parameters

Hugging Face records a total parameter count per repository, and for the four WeMM-Video base checkpoints they run consistently about 1.2 to 1.6 billion above the size in the name.

• WeMM-Video-2B — 3.28B total against a "2B" label.

• WeMM-Video-4B — 5.61B total against a "4B" label.

• WeMM-Video-9B — 10.49B total against a "9B" label.

• WeMM-Video-35BA3B — 36.18B total against a "35B-A3B" label, where only three billion of those are active per token.

Reading the safetensors index of the 4B makes the surplus visible. Of its 1,268 tensors, the language model carries 4.21B parameters, the audio stack carries 1.07B and the vision tower carries 0.33B. Roughly nineteen per cent of the checkpoint is ears. The paper confirms the naming convention is deliberate: the three dense models are built on Qwen3.5-2B, Qwen3.5-4B and Qwen3.5-9B, and the mixture-of-experts model on Qwen3.6-35B-A3B.

That matters for anyone sizing a deployment. A "4B" model that needs room for 5.61B parameters — 10.45 GiB in BF16 before any KV cache or activation — lands in a different box than the number on the card suggests.

Eighteen repos, nineteen minutes, no announcement

The family did not trickle out. Hugging Face creation timestamps put all eighteen repositories — four sizes times five precisions — in a window from 08:53:50 to 08:59:06 UTC on 17 November 2026, roughly fourteen seconds apart, which is the signature of an upload script rather than a person clicking publish. The GitHub repository Tencent/WeMM-Video was created at 02:29:33 UTC that same day, with its substantive initial commit arriving on 19 November and README revisions continuing into 20 November.

A screenshot of the Hugging Face model page for tencent/WeMM-Video-4B showing the '6B params / 10.5 GB' panel, a Safetensors tag, an Image-Text-to-Text pipeline tag, 10.4 GiB of weights, an Apache 2.0 licence, and zero downloads and zero likes.

The paper followed on 19 November: arXiv 2611.26409, "WeMM-Video: Pushing the Frontiers of Open-Source Omni-Modal Video Understanding," twelve authors, submitted at 16:58 UTC. It is a full technical report with seven results tables and a link to an evaluation and reproduction pipeline. So this is not an undocumented drop — it is a documented one that nobody announced. We could not find a T​encent blog post, a WeChat Vision launch page, a press release or a paid tier pointing at the family; the model cards, the GitHub README and the arXiv paper are the entire public record, and they were written for each other.

The reception figures tell the same story. At the time of writing every one of the four base checkpoints showed zero downloads, and the eighteen repositories between them had collected three likes. Neither Artificial Analysis nor LMArena lists a WeMM model. None of that is evidence about quality — it is evidence about distribution, and it is the reason a 4B checkpoint that outscores a 30B on several video benchmarks has no independent number attached to it.

What is inside the checkpoint

Stripping the family framing, the 4B is a standard three-tower omni stack with one unusual backbone choice.

• Language tower — 32 layers, hidden size 2,560, 16 attention heads against 4 key-value heads, head dimension 256, vocabulary 248,320. The backbone is hybrid: one full-attention layer every fourth block, with the other three using linear attention, which is why every serving recipe in the repository carries a float32 SSM cache flag.

• Vision tower — 24 blocks at hidden size 1,024, patch size 16, spatial merge 2, temporal patch 2, projecting to 2,560. It is the backbone's native encoder, kept rather than swapped, and the paper freezes it from the second pre-training stage onward because it "already provides sufficiently strong visual representations."

• Audio tower — Whisper-style, 32 encoder and 32 decoder layers at width 1,280, with a window-level Q-Former compressing 0.1-second windows into 10 audio tokens per second. Audio arrives at 16 kHz in 30-second chunks. This is the part initialised from Qwen2.5-Omni, and at 1.07B parameters it is the single biggest gap between the name and the file.

• Fusion — visual and audio tokens interleave in chronological order with explicit textual timestamps, rather than being pooled into separate streams.

• Modes — an Instruct mode for direct answers and a Thinking mode that emits a reasoning trace before the response, switched at the chat-template level with enable_thinking.

• Licence — Apache 2.0 across the collection, with T​encent's licence file carving out third-party components that "remain subject to their respective original terms." The carve-out is not itemised in the model card.

256K in the config, 8K in the serving command

The text configuration declares max_position_embeddings of 262,144. Every serving command T​encent publishes — both vLLM profiles and both SGLang profiles — passes max-model-len 8192 or context-length 8192. The 256K figure is architectural headroom, not a supported configuration, and nothing in the repository explains how to reach it.

The frame budget has the same shape. The processor's own defaults are generous: 2 frames per second, up to 128 frames, 25,165,824 total pixels. The vLLM recipe overrides all three downward — fps 2, max_frames 16, total_pixels 262144. Sixteen frames is what you get by copying the command. A clip sampled at 16 frames is a different input from a clip sampled at 128, and the paper does not state which sampling its video evaluations used.

• Per-request media limits, as shipped — four images, two videos, two audio items.

• Context, as shipped — 8,192 tokens, with audio consuming 10 tokens per second before any video frames are counted.

For a 4B model those two numbers are the binding constraint on real work. Eight thousand tokens is about thirteen minutes of audio and nothing else; a video with a soundtrack spends the budget on both streams at once. The long-video benchmarks below were run under conditions the repository does not describe.

The benchmarks are the authors' own

Every number in this article comes from the WeMM-Video technical report. The paper says so plainly — "For existing models, we report official results where available and our own evaluations otherwise" — and no third party has reproduced any of it. The mode also moves between tables, which matters more than it usually would because Thinking mode is not free: the video tables report Thinking for the 2B, 4B and 9B, the audio table reports Instruct because the authors judged standard audio benchmarks to be "direct perceptual" tasks, and the image and text tables return to Thinking. Comparing a 4B row in the audio table against a rival row in the video table is not a comparison.

A single-column scoreboard headed 'WeMM-Video-4B - the scoreboard' listing 'Video-MME with audio: 83.9', 'Video-MME v2 with audio: 48.1', 'LVBench video only: 65.3', 'Daily-Omni: 82.5', 'MMAU audio: 75.2' and 'MMMU image: 78.4', with a footer reading 'Vendor-reported; no independent scores yet. Figures from arXiv 2611.26409.'

With that caveat attached, the omni-modal video table is where the 4B earns its file size.

Where the 4B beats a 30B

Table 2 of the paper puts WeMM-Video-4B against Qwen3-Omni-30B-A3B-Instruct and MiniCPM-o-4.5-9B on joint audio-visual benchmarks, all in Thinking mode. The 4B wins most of them, against a model seven times its size.

• Video-MME with audio — WeMM-Video-4B 83.9, Qwen3-Omni-30B-A3B-Instruct 77.3, MiniCPM-o-4.5-9B 69.2.

• Video-MME v2 with audio — 48.1 against 36.3 and 28.1, the widest margin in the table.

• Daily-Omni — 82.5 against 70.7 and 80.2.

• WorldSense — 60.3 against 54.0 and 55.7.

• OmniVideoBench — 49.7 against 45.4 and 37.2.

The silent-video table is stronger still, and it is the one that should decide a purchase. Stripping the soundtrack and testing vision alone: Video-MME 75.7 against 70.5 and 70.4, LVBench 65.3 against 50.2 and 50.9, LongVideoBench 74.8 against 63.5 and 66.0, EgoSchema 74.2 against 66.4 and 60.4. A 15-point lead on LVBench at 4B is not a rounding artefact; long-video comprehension is exactly where small omni models have historically fallen apart.

The image and text rows are competitive rather than dominant. MMMU 78.4 against Qwen3.5-Omni-Flash's 76.9, MathVista 85.2 against 82.9, DocVQA 96.5 against 95.3, IFEval 91.3 against 85.2. Those are the rows a reader is most likely to actually need, and they hold up.

Where it loses, and to what

The losses are specific and worth naming, because a table that only shows the wins is the thing this blog exists to avoid.

• Audio, against a proprietary flash model — Qwen3.5-Omni-Flash scores 80.4 on MMAU and 74.0 on MMAR against WeMM-Video-4B's 75.2 and 67.3. The 4B's one audio win is MMAU-Pro at 68.1 against 64.8.

• Video, in the corners — Video-Holmes 58.8 against MiniCPM-o-4.5-9B's 64.3, and FutureOmni 50.9 against 55.1 for Qwen3-Omni-30B-A3B-Instruct and 56.1 for MiniCPM. Two benchmarks where a smaller or older competitor wins outright.

• Image and text — MMMU-Pro 64.7 against 68.2, OCRBench 85.7 against 89.1, GPQA 73.2 against 76.4, MMLU-Pro 78.3 against 79.9. Between one and six points behind a serving-oriented proprietary model, which for a 5.61B checkpoint is a respectable place to be behind.

The reading is that WeMM-Video-4B was tuned for joint audio-visual video and held its ground everywhere else. If your job is describing what happens in a clip and summarising what is said over it, the numbers are good. If your job is dense OCR or graduate-level text reasoning, the same checkpoint is a compromise.

Running it means patching your own runtime

A screenshot of the Tencent/WeMM-Video GitHub repository showing five stars, zero forks, and the README model-zoo table listing the WeMM-Video 2B, 4B, 9B and 35B-A3B checkpoints against their base models.

Neither vLLM nor SGLang supports this architecture upstream. A search of the vLLM issue tracker turns up no WeMM-Video request, and the repository ships a patch directory instead: source overlays with SHA-256 manifests for four exact upstream releases — vLLM 0.18.0 and 0.31.0, SGLang 0.5.16 and 0.5.19 — each pinned to a specific PyTorch, Transformers and CUDA 12.9 combination, with Python 3.12 and FFmpeg required and TorchCodec needing CUDA 12 NPP on the linker path. SGLang additionally compiles CUDA kernels on first use, so a C++ toolchain and Ninja are part of the install.

The GGUF path is not a shortcut either. It is a three-file bundle — a 4.90 GiB language file, a 0.63 GiB vision projector and a 2.00 GiB audio file — built for a "WeMM native runtime" that you compile from the repository, not for llama.cpp or Ollama. T​encent's own manifest labels the export exported_unvalidated, and the accompanying scope file says quantized_candidate_not_validated. That is unusually candid labelling and it should be read literally: the Q4_K_M bundle is offered, not vouched for.

What T​encent does claim as validated is narrow: patched vLLM loaded the checkpoint on an H20 and completed generation checks for text, image, audio and video with a soundtrack, and the local checkpoint passed a direct Transformers load plus a CPU text-generation check. No throughput, no latency, no memory footprint, no concurrency numbers. For a model whose selling point is video, "it generated" is the floor of what a deployment decision needs.

OrcaRouter does not serve WeMM-Video-4B. Its catalogue lists no WeMM checkpoint at all, so there is no hosted price to quote and no endpoint to point at; today this is a self-host job with a patch applied to somebody else's inference engine. For the vision-only half of the work there are open alternatives on the platform — Qwen3-VL-235B-A22B-Thinking and Qwen3-VL-8B-Thinking both take images and video frames — reachable through one OpenAI-compatible key, with no WeMM equivalent for the audio half. If a compatible host does start serving these weights, provider list price passes through at 0% markup, but that is a condition, not a current fact.

What is still unconfirmed

Four things a buyer would normally get from a launch and cannot get here.

• Price. There is none. No hosted endpoint, no per-token rate, no partner list. The only cost is whatever GPU you already own.

• Independent evaluation. No Artificial Analysis entry, no arena submission, no third-party reproduction. The evaluation pipeline is public, which is the next best thing and is not the same thing.

• Evaluation conditions. Frame counts, pixel budgets and context lengths used for the benchmark runs are not stated in the paper, and the shipped recipe uses 16 frames where the processor defaults to 128.

• The third-party licence carve-out. T​encent states that components inside the collection remain under their original terms but does not enumerate them. The audio stack descends from Qwen2.5-Omni and the backbone from Q​wen3.5, both Apache 2.0, which is reassuring rather than dispositive.

None of that is a reason to dismiss the release. It is a reason to treat the family as what it visibly is: eighteen repositories and a technical report put online in a nineteen-minute window, with the announcement step skipped. The documentation is better than most announced launches. The distribution is zero.

For a team that needs joint audio-visual understanding on hardware they control, WeMM-Video-4B is the most interesting small checkpoint published this month, and the video numbers justify the patching effort. For anyone who needs an endpoint by Friday, the honest answer is that this is not one yet — the weights exist, the runtime is a source overlay, and the only benchmarks are the authors'. Watch for the first independent reproduction, a vLLM upstream merge, or a hosted price. Whichever lands first will tell you more than another row in Table 2.

Compared in this article1

Detected from this article