
WeMM-Video-35B-A3B: Weights Land, Scores Are Tencent's Own
- 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
For three days the Hugging Face repository tencent/WeMM-Video-35BA3B contained one file: a 1.5 KB .gitattributes. The technical report describing WeMM-Video-35B-A3B went up on arXiv on 19 November, calling the checkpoint a new open-source state of the art for joint audio-visual understanding, ahead of Qwen3.5-Omni-Flash and Seed2.0-Mini-0428 and closing on Gemini-3.0-Flash. A reader who believed the report and clicked through to the weights found nothing to download. The 67.39 GiB of BF16 tensors only landed at 03:57 UTC on 20 November — this morning. The flagship was announced a day before it existed, and every number attached to it so far comes from the team that trained it.
That timing is not a scandal, it is a normal research release with the steps out of order. But it does mean the useful question about WeMM-Video-35B-A3B is not "is it good" — the report answers that in the team's own favour, across seven tables — but "what can you verify yourself today, and what are you taking on faith." The repository answers the first half better than most releases do.
The weights arrived this morning, not on the 17th
Hugging Face timestamps make the sequence unusually legible. The nineteen WeMM-Video repositories in Tencent's collection were created inside a six-minute window on 17 November, between 08:53:50 and 08:59:06 UTC — the signature of an upload script rather than a person clicking publish. For the 35B-A3B repository, the commit titled initial commit at 08:53:50 contains exactly one file, and it is not a weight. The commit titled Upload WeMM-Video-35BA3B BF16 release, carrying seventeen safetensors shards, is dated 2026-11-20T03:57:36Z.
The smaller sizes filled in the same morning: 2B at 01:41, 4B at 04:08, 9B at 05:08. The GitHub repository Tencent/WeMM-Video was created on 17 November at 02:29 UTC and its README was still being revised at 05:09 on 20 November. The arXiv report, 2611.26409, was submitted on 19 November at 16:58 UTC.
None of that is unusual for a lab shipping a family. What it changes is the position of anyone reading about the model this week: the report describes a checkpoint that was, for roughly thirty-six hours, a promise. If you read coverage on 19 November and tried to run it, the failure was not your toolchain.

What is inside a 36.18B checkpoint
Hugging Face's own metadata puts the checkpoint at 36,177,564,536 parameters — 36.18B against a "35B-A3B" label. The label is not wrong, it just counts something different from the file size, and the gap between the two is the most practically important thing about this model.
• Language tower — 40 layers at hidden width 2,048, 16 query heads against 2 key-value heads, head dimension 256, vocabulary 248,320. Attention is hybrid: every fourth layer is full attention and the other three are linear attention, which is why every serving command in the repository carries a float32 SSM cache flag.
• Experts — 256 routed experts per layer at intermediate width 512, top-8 routing, plus a 512-wide shared expert. Across 40 layers that is 10,240 routed experts.
• Vision tower — 27 blocks at hidden width 1,152, patch size 16, spatial merge 2, temporal patch 2, projecting to 2,048.
• Audio — Whisper-shaped, 32 encoder and 32 decoder layers at width 1,280, feeding a window-level Q-Former that compresses each 0.1-second window to a single token. That works out to 10 audio tokens per second, from 16 kHz input split into 30-second chunks.
• Provenance — the report says the MoE is built on Qwen3.6-35B-A3B and that the audio encoders are initialized from Qwen2.5-Omni. Video frames carry textual timestamps, and visual and audio tokens are interleaved in chronological order rather than pooled into separate streams.
• Training — approximately 1.2 trillion multimodal tokens through progressive pretraining, then multi-teacher on-policy distillation, an anti-repetition preference stage the authors call MrDPO, and specialist reinforcement learning.
• Licence — Apache 2.0 across the collection, with a carve-out for third-party components that "remain subject to their respective original terms." The carve-out is stated, not itemised.
Four quantised variants are published next to the BF16 checkpoint, and their manifest files are worth reading because of how frankly they are labelled. AWQ W4A16 totals 22.69 GiB and its manifest says quantized_unvalidated. NVFP4 W4A16 is a single 24.11 GiB safetensors file whose manifest claims only cpu_NVFP4_weight_artifact_independently_reread — the weights were re-read on CPU, which is a checksum, not an evaluation. The GGUF Q4_K_M bundle is three files (22.61 GiB language, 0.84 GiB vision projector, 1.99 GiB audio) and its manifest says exported_unvalidated, with the per-tensor scope file saying quantized_candidate_not_validated.
The fifth variant is missing. The model card and the GitHub model-zoo table both link an FP8 Block128 repository; the Hub returns 401 for it, the same answer it gives for a repository name that has never existed, while every live sibling returns 200. The Hugging Face collection has a hole at position 1, exactly where that repository was meant to sit. This is a release that is still being assembled as you read it.
One measured number, and it is audio
The model card reports exactly one evaluation the vendor ran itself: MMAR1000 at 72.7%, 727 of 1,000. The conditions are given in full, which is the good part — audio only, full audio decoded at 16 kHz, Instruct mode, greedy decoding, at most 256 output tokens, patched vLLM 0.18 on two H20 GPUs with tensor parallelism.
Everything else on that card is a statement that something loaded. Patched vLLM loaded the checkpoint on H20 and completed generation checks for text, image, audio and video with a soundtrack; the local checkpoint passed a direct Transformers load and a CPU text-generation check. That is the floor of what a deployment decision needs, and Tencent does not pretend otherwise.
There is one loose thread worth pulling. The serving guide in the same GitHub repository carries a validation table whose last row reads "GPU generation after downloading a checkpoint from the Hub is pending." The model card says the H20 run, which loads from the Hub, completed generation for all four modalities. Two documents written by the same team disagree about how much has been demonstrated. The card's claim is the more specific and probably the more current; the disagreement is a fair reason to budget a day of your own smoke tests rather than a week of integration.
The report's audio table also lists 73.6 on MMAR for this checkpoint, which is a different harness from the card's MMAR1000 run and should not be subtracted from 72.7 to produce a "real" number. The only honest reading is that Tencent's own audio figure lands in the low seventies and no second party has measured it.
28 of 30 against Qwen's omni flash
The report's Table 1 is the flagship's case, and it is a large one. Across the 30 single-valued benchmarks in that table — joint audio-visual tasks, audio-visual hallucination, music understanding, interaction, captioning — WeMM-Video-35B-A3B in Thinking mode is ahead of Qwen3.5-Omni-Flash on 28 and behind on 2. Against Seed2.0-Mini-0428 it is ahead on 26 and behind on 4. Against Gemini-3.0-Flash it is ahead on 9 and behind on 21.
The wins that matter most are not the tidy ones:
• Video-MME with audio — 86.4 for WeMM-Video-35B-A3B, 79.3 for Qwen3.5-Omni-Flash, 83.6 for Seed2.0-Mini-0428, 88.6 for Gemini-3.0-Flash.
• Video-MME v2 with audio — 54.9 against 36.6 and 51.7, with Gemini-3.0-Flash at 63.2. An eighteen-point lead over the closest open competitor on the harder split of the same benchmark.
• AV-SpeakerBench — 72.6 against 65.2 and 68.5, still behind Gemini at 75.9. Telling apart who spoke, over video, is the task that separates real audio-visual fusion from a vision model with a transcript bolted on.
• SciVideoBench and EvaOmni — 69.4 against 51.3, and 57.8 against 30.4. Both are reasoning-heavy; both are where a weaker omni model tends to collapse.
The losses are equally specific and worth naming, because a table that only shows wins is the thing this blog exists to avoid:
• JointAVBench — 65.9, behind Qwen3.5-Omni-Flash at 69.0 and Gemini-3.0-Flash at 68.4, ahead only of Seed2.0-Mini-0428 at 64.7.
• MUSIC-AVQA — 79.5 against Qwen's 82.1, though ahead of Gemini's 76.2. Music is where Qwen's audio stack still has the edge.
• Against Seed2.0-Mini-0428 — OmniCompass-AV 46.3 against 47.9, OmniVChat 48.9 against 49.8, O-IN-Bench 49.0 against 49.4, video-SALMONN 2's testset 28.3 against 32.7. Three of those four are within a point.
• Against Gemini-3.0-Flash — the gap is widest on long and future-oriented video: MMOU 74.5 against 84.7, FutureOmni 56.2 against 65.3, OmniVideoBench 55.3 against 63.3, LVOmniBench 52.6 against 59.8. "Narrows the gap" is a fair description; "closes it" would not be.
Every number above is Tencent's. The report says so itself — "For existing models, we report official results where available and our own evaluations otherwise" — and moves between Instruct and Thinking mode depending on the table, which it also states. No third party has reproduced any of it, and neither Artificial Analysis nor LMArena lists a WeMM model.

Adding ears barely dented the vision scores
The claim in this release that deserves more attention than the headline benchmark is a quieter one: that bolting an audio stack onto a vision-language model did not cost it much vision. Omni-modal training usually does cost something — the model spreads capacity across modalities and gives up ground on silent video.
The report tests that directly by putting WeMM-Video-35B-A3B against its own backbone, Qwen3.6-35B-A3B, on 31 silent-video benchmarks. The flagship is ahead on 20 and behind on 11, and most of the eleven are inside a point.
• Where it gained — TimeLens-Charades 54.8 against 47.4, F-16-NBA 89.9 against 82.8, CrossVid 50.2 against 45.2, OVO-Bench 63.9 against 61.0, TemporalBench 45.9 against 43.3.
• Where it lost — EgoSchema 74.2 against 77.2, Video-MME 80.4 against 82.5, MVBench 73.1 against 74.6, MLVU 85.3 against 86.2. The rest are fractions of a point.
Read the pattern rather than the count: the gains cluster in temporal grounding and long-video association, the losses in first-person egocentric video. That is a plausible trade for a model tuned on video with sound. Against Qwen3.5-Omni-Flash on the same silent-video table it is ahead on 30 of 31, losing only VideoVista, 83.5 to 84.1.
The image and text tables show the same shape. Against Qwen3.5-Omni-Flash, 26 of 28 image benchmarks go to WeMM, with RefCOCO (89.5 against 92.6) and TallyQA (80.7 against 81.2) the exceptions. Against its own backbone the image table is a near-tie — 13 ahead, 14 behind — and the text table is where the cost shows up: MMLU-Pro 84.7 against 85.2, GPQA 83.8 against 86.0, IMOAnswerBench 74.8 against 78.9, SuperGPQA 63.4 against 64.7. A few points of hard reasoning given up for a working pair of ears is a normal and defensible trade, but it is a trade, and anyone routing graduate-level text through this checkpoint should know it exists.
36B of memory, 3B of arithmetic
Here is the arithmetic that decides whether this model is deployable for you, and it comes out of the config rather than the prose. Of the 36.18B parameters, roughly 32.2B sit in routed experts: 40 layers times 256 experts times three projections at width 2,048 by 512. Top-8 routing touches one thirty-second of that per token, about 1.0B parameters, and the label's "3B active" is consistent once attention, the shared expert and the embedding table are counted.
So the compute bill is small and the memory bill is not:
• BF16 — 67.39 GiB across seventeen shards. Tencent's own validation used tensor parallelism across two H20s. There is no published tensor-parallel number for larger configurations; the serving guide says only "plan memory for the full expert weights and set --tensor-parallel-size as needed."
• NVFP4 W4A16 — 24.11 GiB, which fits inside a single 48 GB card with room for cache. The recipe is a separate patched vLLM 0.31.0 environment with ModelOpt mixed precision and Marlin, and it runs with --max-num-seqs 1.
• AWQ W4A16 — 22.69 GiB on patched vLLM 0.18.0 using the checkpoint's own compressed-tensors configuration.
• GGUF Q4_K_M — a 24.6 GiB three-file bundle, but for Tencent's own native runtime, not llama.cpp or Ollama. Its manifest advertises it as exported and unvalidated.
Nothing here runs on an unpatched engine. The repository ships a patch directory: source overlays with SHA-256 manifests for exactly four 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, TorchCodec needing CUDA 12 NPP on the linker path, and SGLang compiling CUDA kernels on first use. A search of the vLLM issue tracker turns up no upstream request for this architecture. This is a self-host job with an overlay applied to somebody else's inference engine.

Two configuration details bound what you can actually feed it. The text config declares 262,144 positions, but the card's own vLLM command passes --max-model-len 8192, and every SGLang recipe caps context at 8,192 as well — the long figure is headroom, not a supported setting. Frame sampling is the same story: the processor defaults to 2 fps and up to 128 frames, while the published recipe overrides to 16 frames and a 262,144-pixel budget. At 10 audio tokens per second, an 8,192-token context is a little over thirteen minutes of audio and nothing else.
OrcaRouter does not serve WeMM-Video-35B-A3B, and the catalogue lists no WeMM checkpoint at all, so there is no hosted price to quote. What we do serve is the model this one was built on: Qwen3.6-35B-A3B is on the platform at $0.25 per million input tokens and $1.48 per million output over a 262,144-token context, with text, image and video input. If the half of this job you need is "watch the clip and reason about it" rather than "watch the clip and transcribe the room," that is a hosted stand-in available this afternoon instead of a 67 GiB download and a source-level patch. Provider list price passes through at 0% markup, and every model sits behind one OpenAI-compatible key with automatic failover, so trying a second opinion costs a config change rather than a second contract.
What nobody outside WeChat has checked
Four things a buyer would normally get from a launch and cannot get here.
• Independent evaluation. None. The evaluation and reproduction 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 report, and the shipped recipe samples 16 frames where the processor defaults to 128. A 16-frame sample of a long clip is a different input from a 128-frame sample of the same clip.
• The deployment claim. The report states the family is "deployed across multiple core video services within WeChat," naming live-stream commerce, short-form drama analysis and interactive video. That is a serious claim if true — production traffic is the one evaluation nobody can argue with — and it is completely unverifiable from outside. No latency, throughput, concurrency or memory figures are published anywhere in the release.
• The quantised artefacts. Three of the four published variants describe themselves, in their own manifests, as unvalidated. AWQ, GGUF and NVFP4 are offered, not vouched for.
None of that is a reason to dismiss the release. It is a reason to read it as what it visibly is: a well-documented research release whose flagship weights are hours old, whose paper arrived first, whose FP8 repository has not been created yet, and whose benchmark lead over the best open omni competitor is large enough to be worth reproducing rather than doubting.
What would settle it is the same list it always is. A vLLM upstream merge would end the patching. A first independent reproduction on Video-MME v2 with audio would confirm or puncture the eighteen-point gap. A hosted price, or the missing FP8 repository, would tell you whether Tencent intends this to be used or merely cited. Until one of those lands, the honest summary is that WeMM-Video-35B-A3B is the most interesting open omni-modal checkpoint of the month, that its vision scores appear to survive the addition of hearing, and that all of its numbers currently come from the people who trained it.
