On November 16, 2026, Cohere announced Video Intelligence in Compass, currently available in Private Preview while recruiting design partners rather than open for general self-serve availability. The release brings semantic moment retrieval into Compass enterprise search and agent workflows.

Timestamped Moment Retrieval, Unified Governance & Design Partners

Query-Time Timestamped Moments & Compass MCP: Understanding both spoken audio and on-screen visuals, Video Intelligence determines relevant start and end times dynamically at search time rather than returning hour-long recordings or relying on fixed chunks. Each result links back to the source footage with timestamps, transcripts, and visual descriptions, and agents can connect to the Compass MCP server to reason over slides, charts, and demos.

One Governed Index & Design Partner Deployment: Video is indexed alongside enterprise documents with permission-aware retrieval across meetings, training, media archives, and legal compliance workflows. Design partners can deploy in Cohere's managed cloud or their own environment to evaluate real-world footage.

Figure 2 Cohere illustrative architecture diagram showing unified, permission-governed indexing across video and enterprise documents
Figure 2 Cohere illustrative architecture diagram showing unified, permission-governed indexing across video and enterprise documents

Image source: Cohere Blog (https://cohere.com/blog/video-intelligence-preview).

Cohere Vendor Evaluation Results (nDCG@10)

Average 0.686 nDCG@10 Across 11 Datasets: In Cohere's benchmark across 11 datasets (six public benchmarks and five internal enterprise datasets, totaling 11,133 queries), Cohere Video Intelligence achieved an average nDCG@10 of 0.686—4.4 points ahead of Qwen3-VL-Embedding-8B (0.642) and ahead of Gemini Embedding 2 (0.633), Qwen3-VL-Embedding-2B (0.590), EmbeddingGemma 2 (0.556), TwelveLabs Marengo 3.5 (0.545), Amazon Nova Multimodal Embeddings (0.474), and Jina Embeddings v5 Omni Small (0.398).

Domain Leads & Per-Dataset Nuances: Cohere led across all three domain averages: Public benchmarks (0.591 vs. Amazon Nova's 0.567, +2.4 points), Enterprise knowledge (0.872 vs. Qwen3-VL-Embedding-8B's 0.852, +2.0 points), and Sports (stated in text as 5.1 points ahead of Gemini Embedding 2, with chart values shown as 0.695 vs. 0.645). Individual datasets show mixed leaders, with competitors outscoring Cohere on MSR-VTT (Qwen3-VL-Embedding-8B at 0.701 vs. 0.567), QVHighlights (TwelveLabs Marengo 3.5 at 0.523 vs. 0.511), and OSHA training (Qwen3-VL-Embedding-8B at 0.912 vs. 0.831).

Figure 1 Cohere's vendor evaluation chart comparing Video Intelligence against seven video embedding models in nDCG@10 across 11 datasets (11,133 queries)
Figure 1 Cohere's vendor evaluation chart comparing Video Intelligence against seven video embedding models in nDCG@10 across 11 datasets (11,133 queries)

Image source: Cohere Blog (https://cohere.com/blog/video-intelligence-preview).