
WeMM-Video-9B: Why Its 4-Bit Builds Are Still 13 GiB
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 182 tok/s
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 356 tok/s
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 87 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 147 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-0737 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 92 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 135 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 81 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 117 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 110 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 66 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 91 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 42 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 211 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 109 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash 09102026-10-17$0.15 / $0.60 per 1M tokens · 160 tok/s
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
A four-bit quantisation should land at roughly a quarter of the weight file it came from. WeMM-Video-9B is 19.54 GiB in BF16, and its AWQ W4A16 build is 12.86 GiB. That is a 34% cut, not a 75% one — and the FP8, AWQ, NVFP4 and GGUF builds all land within 0.42 GiB of each other, so the choice between them buys you almost no memory at all. The exporters' own manifests explain why: the 4-bit pipelines compress only the dense feed-forward matrices — 4.83 billion parameters out of 10.49 billion — while the vision tower, audio tower, embedding, output head and attention weights all keep their original precision, and 5.7 billion parameters at two bytes each set a floor no W4 recipe can go under. Tencent published the weights, the quantisation recipes and a 12-author technical report on 17–19 November without a launch post, and the part nobody has written up yet is the arithmetic — which also happens to be the part that decides whether this model fits on the GPU you already own. On Tencent's own tables it beats Qwen3-Omni-30B-A3B-Instruct on 35 of the 37 joint audio-visual figures the two share; on whether its four-bit builds are safe to deploy, the record is thinner than nineteen repositories suggest.
A 10.49B stack wearing a 9B label
The repository tencent/WeMM-Video-9B was created at 08:55:54 UTC on 17 November 2026, fourteen seconds after its AWQ sibling and eleven minutes after the 35B-A3B flagbearer. Nineteen WeMM-Video repositories exist under the Tencent organisation, all created inside a five-minute window that morning, which is the signature of an upload script rather than a person publishing one model at a time. The GitHub project Tencent/WeMM-Video was created six hours earlier, at 02:29 UTC, its substantive initial commit landed on 19 November, and README revisions were still arriving on 20 November. The paper, arXiv 2611.26409, was submitted at 16:58 UTC on 19 November.

None of that was announced. We could not find a Tencent blog post, a WeChat Vision launch page, a press release or a paid tier pointing at the family. At the time of writing all nineteen repositories showed zero downloads and, between them, three likes. Artificial Analysis lists no WeMM model. Neither of those facts says anything about quality — they say the distribution step was skipped.
The "9B" in the name is the language backbone, not the checkpoint. Hugging Face records 10,488,603,384 parameters across 1,305 tensors, and reading the five safetensors shards by name puts them in four piles:
• Language model — 7.94B parameters, including a 1.02B token embedding.
• Output head — 1.02B parameters. The embeddings are untied, so this is a second full copy of the 248,320-token vocabulary at 4,096 hidden.
• Audio tower — 1.08B parameters. A Whisper-style encoder and decoder, 32 layers each at width 1,280, feeding a two-layer Q-Former that compresses 0.1-second windows into 10 audio tokens per second.
• Vision tower — 0.46B parameters. 27 blocks at hidden size 1,152, patch size 16, spatial merge 2, temporal patch 2.
So roughly a fifth of the file is ears and eyes. The paper is explicit about provenance: the dense 2B, 4B and 9B are built on Qwen3.5-2B, Qwen3.5-4B and Qwen3.5-9B, the mixture-of-experts model on Qwen3.6-35B-A3B, the audio encoders are initialised from Qwen2.5-Omni, and the native visual encoders are kept rather than swapped and frozen from the second pre-training stage onward because they "already provide sufficiently strong visual representations."
The backbone itself is a hybrid: 32 layers, hidden size 4,096, 16 attention heads against 4 key-value heads at head dimension 256, with one full-attention layer after every three linear-attention layers and a float32 SSM cache — which is why every serving command in the repository carries a mamba-ssm-cache-dtype float32 flag. Position embeddings are declared at 262,144 tokens; interleaved multimodal RoPE sections are 11/11/10. Licence is Apache 2.0 with a third-party carve-out that is not itemised on the model card.
Four builds, four near-identical sizes
Here is the whole quantisation ladder, measured from each repository's own file listing rather than from a rounded figure on a card:
• BF16 base — 19.54 GiB across five shards.
• FP8 Block128 — 13.10 GiB, one file. Needs an FP8-capable GPU and patched vLLM 0.18.0.
• AWQ W4A16 — 12.86 GiB, group-128 INT4 in compressed-tensors format. Patched vLLM 0.18.0.
• NVFP4 W4A16 — 13.07 GiB, ModelOpt 0.45.0 mixed precision with group size 16. Patched vLLM 0.31.0 with Marlin, SM120 reference configuration.
• GGUF Q4_K_M — 13.28 GiB as a three-file bundle: 10.41 GiB language, 2.01 GiB audio, 0.86 GiB vision projector.

The spread across those four compressed builds is 0.42 GiB, which is smaller than the buffer you would leave for a driver update. If you were choosing between them to fit a card, you are choosing the wrong variable: pick on which GPU you own and which runtime you can patch, because the memory difference is not there.
The naive expectation — a quarter of 19.54 GiB, so about 4.9 GiB — is off by a factor of two and a half. Even the headroom against BF16 is modest: 6.68 GiB saved by AWQ, which is a real 34%, but nothing like the "four times smaller" that a W4 label implies.
What the exporters protect, and why
Both the AWQ and GGUF manifests state exactly which tensors they were allowed to touch, and the answer explains the size plateau.
• Only the dense feed-forward matrices are quantised. The AWQ manifest records 96 quantised language matrices — three per layer across 32 layers, each 4,096 by 12,288. That is 4.83B parameters out of 10.49B, or 46% of the checkpoint. Everything else keeps its precision.
• The attention path is restored to BF16. The AWQ manifest carries an attention_restoration block listing 152 attention projection matrices and 32 input norms copied back from the source checkpoint with their SHA-256 hashes. The GGUF scope file reports the same count — protected_attention_BF16: 152 — and adds that the embedding and output tensors are the only two other protected items.
• The frontend and both media towers never enter the quantiser. 881 tensors are listed as protected frontend, embedding and head tensors in the AWQ manifest — 333 from the visual tower, 545 from the audio tower, the token embedding, the untied output head and the final norm. At BF16 that is about 3.57B parameters, or 7.1 GB of the 21.0 GB file, sitting outside the compression step entirely.
The consequence is the number a deployment plan actually needs: budget 13–13.5 GiB of VRAM for weights whichever build you pick, then add the KV cache and the float32 SSM cache on top. A "4-bit 9B" that needs 13 GiB of weights is not a problem with the model — the vocabulary alone is 248,320 tokens embedded twice, and the audio tower is a 1.08B-parameter Whisper sitting in the middle of it — but it is the difference between fitting on a 16 GB card with room to work and not fitting at all.
Validated to load, not validated to be equal
Every quantisation card in this family reports a smoke test, and every manifest reports something blunter. Read the two side by side, because the gap between them is the whole risk of using a compressed build.
• FP8 Block128 — card: patched vLLM loaded it on an H20 and completed text, image, audio and video-with-audio generation checks, and a runtime precision audit confirmed the W8A8 path. Manifest: status quantized_and_independently_reloaded_native_inference_pending, with native_inference_validated: false and native_low_precision_gemm_validated: false recorded as separate flags.
• AWQ W4A16 — card: patched vLLM loaded it on an H20, completed the same four checks, and the precision audit confirmed INT4 feed-forward weights against BF16 attention. Manifest: status quantized_unvalidated, with a diagnostic_variant naming a BF16-attention reference build alongside the INT4 one.
• NVFP4 W4A16 — card: patched vLLM 0.31.0 loaded and ran "the PRO 5000 reference weights" and passed five generation checks, including silent video. Manifest: the same three claims are recorded as flags and every one of them reads false — native_FP4_GEMM_validated, PRO5000_inference_validated, vllm_validated — under a status of cpu_NVFP4_weight_artifact_independently_reread. What the manifest does attest is a numeric check: 96 matrices and 4,831,838,208 elements dequantised independently, mean relative RMSE 0.095 against a 0.35 limit, with 1,209 protected tensors confirmed bit-exact. That element count is the same 4.83 billion the scope statement names, which is how you know the arithmetic in this piece and the exporter's agree.
• GGUF Q4_K_M — card: the native CPU runtime loaded the bundle and generated text, image, audio and video output. Manifest status exported_unvalidated; scope status quantized_candidate_not_validated.
"It loaded and it generated" is a real result and worth having — plenty of quantised multimodal checkpoints fail at exactly that step. It is not the result that decides a deployment, because none of these cards reports a single accuracy number for the compressed build against its own BF16 source. There is no measured delta on MMAU, on Video-MME, on transcription WER. The one precision claim made anywhere is structural, from the AWQ audit: INT4 in the feed-forward, BF16 in attention. That describes the file. It does not tell you what the compression cost, and on a model whose margin over Qwen3-Omni-30B-A3B-Instruct on Video-MME with audio is 7.8 points the size of that cost is not a rounding question.
So the practical rule for this family is to re-run one benchmark you care about on the exact build you intend to ship, against the BF16 base, before you trust a table from the paper. That is a few hours of work and it is the only way to know whether the INT4 feed-forward can be swapped in under the video results, or whether the 34% saving is being paid for in the numbers the card is selling.

Two mechanical traps sit under that.
The first is the GGUF. "GGUF Q4_K_M" reads as "runs in llama.cpp or Ollama." It does not. The bundle is built for Tencent's own native runtime, which you compile from the repository; the model card's quick start is install-cpu/bin and a native/run_hf.py launcher, and the serving guide routes GGUF to that runtime and to nothing else. If your pipeline is llama.cpp-shaped, this checkpoint is not the shortcut it looks like.
The second is a documentation flaw worth naming because it will confuse anyone reading the manifests in order. All three FP8 manifests — for the 2B, the 4B and the 9B — carry the same official_repo field: Qwen/Qwen3.5-35B-A3B-FP8. It is a copy-pasted provenance string in the recipe record, not a statement that a 9B checkpoint is secretly a 35B mixture-of-experts. Nothing else in the file supports that reading: the same manifest's source list points at tencent/WeMM-Video-9B's five shards with matching hashes, and the architecture block records a dense 32-layer model with zero experts. The 35B-A3B is a real separate model; it is also the one size in the family with no FP8 repository of its own, having base, AWQ, GGUF and NVFP4 only.
The 9B against a 30B, on Tencent's tables
Every benchmark number below is the authors' own, from arXiv 2611.26409, and no third party has reproduced any of it. The paper says as much: existing models are reported at "official results where available and our own evaluations otherwise." Treat this section as a description of a claim, not a measurement.
On the joint audio-visual table, in Thinking mode, the 9B beats Qwen3-Omni-30B-A3B-Instruct on 35 of the 37 figures the two columns share, tying one and losing one — against a model roughly three times its size:
• Video-MME with audio — 85.1 against 77.3. MiniCPM-o-4.5-9B sits at 69.2.
• Video-MME-v2 with audio — 51.6 against 36.3, a 15.3-point gap.
• Daily-Omni — 83.6 against 70.7, though MiniCPM-o-4.5-9B's 80.2 makes this a three-way result rather than a walkover.
• SciVideoBench and MMOU — 64.2 against 49.9, and 71.7 against 60.5.
• EvaOmni and OmniReasoningBench — 50.7 against 26.1, and 43.9 against 24.1.
• The one loss — FutureOmni, 53.7 against 55.1. MiniCPM-o-4.5-9B scores 56.1 there.
Silent video is the more useful table, because it contains a same-family control: Qwen3.5-9B, the backbone this checkpoint is built on. Read the two columns side by side and the post-training has a shape. It gains on temporal grounding and temporal perception — F-16-NBA 88.7 against 75.8, TimeLens-QVHighlights 66.8 against 64.3, TimeLens-ActivityNet 55.3 against 52.2, TOMATO 48.5 against 45.0 — and loses on long-video summarisation and general recognition: MLVU 82.4 against 84.4, LVBench 68.2 against 70.0, VideoVista 82.2 against 84.7, VSI-Bench 54.5 against 58.3, MVBench 71.4 against 74.4. So this is not a uniform upgrade over the backbone it inherits. It is a model tuned toward when and where something happened, at a measurable cost in the aggregate video scores and spatial reasoning the backbone already did better.
Audio is where the 9B stops winning. In Instruct mode the 9B scores 70.9 on MMAU-Pro against Qwen3.5-Omni-Flash's 64.8 and 76.8 on MMSU against 72.2 — but loses MMAU 77.0 to 80.4 and MMAR 72.8 to 74.0, and loses speech recognition outright: LibriSpeech word error rate 1.5 and 3.1 against 1.3 and 2.4, WenetSpeech character error rate 5.4 and 6.8 against 4.4 and 5.5. Text and knowledge tell the same story: MMLU-Pro 81.6 against the Qwen3.5-9B backbone's 82.5, GPQA 77.8 against 81.7, HMMT 75.0 against 82.9. If the transcript is the product, a dedicated recogniser still wins.
One comparison in the paper should be read with care. The omni-modal table runs the WeMM models in Thinking mode while the audio table runs them in Instruct, because the authors judged the standard audio benchmarks to be direct perceptual tasks. That is a defensible split, and it also means a WeMM row in one table is not comparable to a rival row in the other.
Eight seconds of video, by default
The gap between what the checkpoint can accept and what the shipped recipes pass it is large enough to change results, so it belongs in any evaluation plan.
• Frame budget — the processor defaults to 2 fps, a minimum of 4 frames, a maximum of 128 frames and 100,663,296 total pixels. The vLLM recipe overrides all four to 2 fps, 16 frames and 262,144 pixels. At 2 fps, sixteen frames is eight seconds of coverage from the middle of a clip.
• Context — the text config declares 262,144 position embeddings. Every serving command Tencent publishes passes max-model-len 8192 or context-length 8192, and nothing in the repository explains how to reach the larger figure.
• Audio cost — 10 tokens per second, so an 8,192-token budget is about thirteen minutes of audio and nothing else. A video with its soundtrack spends that budget on frames and audio at once.
• Per-request media — four images, two videos, two audio items, per the shipped limits.
Nobody outside WeChat Vision has published which sampling the benchmark runs used. If you are reproducing a number from the table above, that is the first unknown to pin down, and the shipped 16-frame default is the wrong place to start.
What you can actually do with it today
OrcaRouter does not serve WeMM-Video-9B, and its catalogue lists no WeMM checkpoint at all. There is no hosted price to quote and no endpoint to call, so today this is a self-host job: patch vLLM or SGLang from the repository's overlay directory, pin the exact PyTorch, Transformers and CUDA 12.9 combination the patch manifest names for your engine version, and budget the 13 GiB plus cache. The GGUF route needs Tencent's native runtime built from source.
For the vision half of the work there are open alternatives already 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 alongside 200+ other models, at 0% markup over provider list price. Neither of them hears anything, and there is no WeMM equivalent for the audio half. If you want to try a self-hosted 9B without putting a production path on it, routing it behind a failover rule is the cheap way to find out whether the patched runtime holds at your concurrency before you commit a card to it.
The four-bit builds, meanwhile, are the right thing to wait on rather than the right thing to adopt. Their own manifests call themselves unvalidated, and the size arithmetic says the compression they offer is a 34% saving on weights — worth having, not worth re-architecting a deployment for. The signal to watch is not another quantisation repository. It is the first manifest in this family that drops the word pending, followed by the first independent reproduction of a table that currently exists in exactly one place: the authors' PDF.
Compared in this article1
Detected from this article
