Video understanding, particularly with accompanying audio, remains a central challenge in multimodal intelligence. In this report, we introduce WeMM-Video, a family of Chinese-English bilingual omni-modal foundation models spanning 2B, 4B, 9B, and 35B-A3B variants. Centering on joint audio-visual perception and reasoning, WeMM-Video unifies text, image, audio, and video processing within a hybrid-thinking paradigm, seamlessly supporting both a non-thinking instruct mode for fast, direct responses and a thinking mode for explicit step-by-step reasoning. The models are trained on a large-scale multimodal corpus of approximately 1.2 trillion tokens through progressive pre-training and a multi-stage post-training pipeline incorporating multi-teacher OPD, anti-repetition MrDPO, and specialist RL. Evaluated across more than 120 benchmarks, WeMM-Video demonstrates superior performance in omni-modal video understanding and reasoning, while maintaining highly competitive capabilities across general audio, silent video, image, and text tasks. Specifically, in omni-modal video evaluations, our flagship WeMM-Video-35B-A3B establishes a new open-source state of the art, substantially outperforming proprietary models such as Qwen3.5-Omni-Flash and Seed2.0-Mini-0428 while narrowing the gap with the strong Gemini-3.0-Flash, and our compact WeMM-Video-9B surpasses the larger open-source Qwen3-Omni-30B-A3B-Instruct. We publicly release the WeMM-Video family along with our evaluation and reproduction pipeline to foster further research and development in the open-source community.
Figures & tables
Figure 1: WeMM-Video can perceive and understand omni-modal videos, perform complex multimodal reasoning, and effectively handle general tasks across audio, image, and text modalities.
Figure 2: Model architecture of WeMM-Video.
Figure 3: Composition of the WeMM-Video training data.
Benchmark 1
Gemini-3.0-
Seed2.0-Mini-
Qwen3.5-Omni-
WeMM-Video-35B-A3B
Flash
0428
Flash
(Instruct)
(Thinking)
General Audio-Visual Understanding
Video-MME w/ audio
88.6
83.6
79.3
84.2
86.4
Video-MME-v2 w/ audio
63.2
51.7
36.6
47.5
54.9
Daily-Omni
82.4
82.5
81.8
81.5
83.4
WorldSense
63.8
61.3
57.9
59.6
62.7
Table 1: Comparison of WeMM-Video-35B-A3B with leading proprietary omni-modal models on joint audio-visual video benchmarks. Best results excluding the gray-shaded reference column are in bold.
Benchmark 1
Qwen3-Omni-
MiniCPM-o
WeMM-Video-
30B-A3B-Instruct
4.5-9B
2B (Thinking)
4B (Thinking)
9B (Thinking)
General Audio-Visual Understanding
Video-MME w/ audio
77.3
69.2
72.4
83.9
85.1
Video-MME-v2 w/ audio
36.3
28.1
37.7
48.1
51.6
Daily-Omni
70.7
80.2
76.5
82.5
83.6
WorldSense
54.0
55.7
52.4
60.3
63.4
Table 2: Comparison of WeMM-Video-2B, 4B, and 9B with open-source omni-modal baselines on joint audio-visual video benchmarks. Best results are in bold.
Benchmark 1
Gemini-3.0-
Seed2.0-Mini-
Qwen3.5-Omni-
Qwen3.6-
WeMM-Video-35B-A3B
Flash
0428
Flash
35B-A3B
(Instruct)
(Thinking)
General Video Understanding
Video-MME
86.4
78.1
77.0
82.5
78.0
80.4
Video-MME-v2
55.6
44.4
33.7
45.3
39.8
45.2
MVBench
70.0
73.4
70.8
74.6
72.8
73.1
Video-MMMU
86.9
81.0
64.0
83.7
80.8
83.6
Table 3: Comparison of WeMM-Video-35B-A3B with proprietary and open-source baselines on silent video benchmarks. Best results excluding the gray-shaded reference column are in bold.
Benchmark 1
Qwen3-Omni-
Qwen3.5-
MiniCPM-o-
WeMM-Video-
30B-A3B-Instruct
9B
4.5-9B
2B (Thinking)
4B (Thinking)
9B (Thinking)
General Video Understanding
Video-MME
70.5
78.4
70.4
67.6
75.7
78.5
Video-MME-v2
29.7
42.4
26.8
32.8
39.5
40.6
MVBench
68.8
74.4
63.4
64.6
70.0
71.4
Video-MMMU
54.7
78.9
52.9
65.1
79.6
79.2
Table 4: Comparison of WeMM-Video-2B, 4B, and 9B with open-source baselines on silent video benchmarks. Best results are in bold.
Benchmark
Qwen3.5-Omni-
Qwen3-Omni-
MiniCPM-o
WeMM-Video-
Flash
30B-A3B-Instruct
4.5-9B
2B (Instruct)
4B (Instruct)
9B (Instruct)
35B-A3B (Instruct)
General Audio Understanding
MMAU
80.4
77.5
76.9
72.4
75.2
77.0
78.0
MMAU-Pro
64.8
60.2
62.7
62.9
68.1
70.9
70.6
MMAR
74.0
73.2
66.3
56.9
67.3
72.8
73.6
MMSU
72.2
69.0
67.4
65.6
72.6
76.8
76.0
Table 5: Evaluation results of the WeMM-Video family and omni-modal baselines on standalone audio benchmarks. Best results are in bold.
Benchmark
Qwen3.5-Omni-
Qwen3.6-
Qwen3.5-
WeMM-Video-
Flash
35B-A3B
9B
2B (Thinking)
4B (Thinking)
9B (Thinking)
35B-A3B (Thinking)
STEM/Puzzle
MMMU
76.9
81.7
78.4
61.2
78.4
78.9
81.1
MMMU-Pro
68.2
75.3
70.1
49.4
64.7
69.4
72.9
MathVision
65.4
80.4
78.9
49.2
73.3
76.0
79.6
MathVista (mini)
82.9
86.4
85.7
78.8
85.2
86.1
86.6
Table 6: Evaluation results of the WeMM-Video family and baselines on image understanding benchmarks. Best results are in bold.
Benchmark
Qwen3.5-Omni-
Qwen3.6-
Qwen3.5-
WeMM-Video-
Flash
35B-A3B
9B
2B (Thinking)
4B (Thinking)
9B (Thinking)
35B-A3B (Thinking)
Knowledge
MMLU-Pro
79.9
85.2
82.5
63.7
78.3
81.6
84.7
MMLU-Redux
90.0
93.3
91.1
75.7
88.7
90.5
93.7
SuperGPQA
54.9
64.7
58.2
36.3
52.4
56.7
63.4
C-Eval
86.0
90.0
88.2
72.0
84.8
87.6
90.4
Table 7: Evaluation results of the WeMM-Video family and baselines on pure text benchmarks. Best results are in bold.
School of Intelligence Science and Technology, Peking University · State Key Laboratory of General Artificial Intelligence, Peking University · Alibaba Token Hub, Alibaba Group