Cohere releases Embed 5 Pro and Fast: multimodal embeddings that share one vector space, top ViDoRe V3
On Sept 30, 2026 Cohere released Embed 5 in two tiers, Pro (embed-v5.0-pro) and Fast (embed-v5.0-fast). Both take text, images and mixed inputs in 100+ languages with a 128K-token context and share one embedding space, so teams can index with Pro and query with Fast without re-indexing. Cohere reports Pro at 85.8 on ViDoRe V3, ahead of Voyage 4 Large (83.7) and Gemini Embedding 2 (83.2). Text pricing is $0.12 (Pro) and $0.08 (Fast) per 1M tokens.
Key facts
- API ids: embed-v5.0-pro and embed-v5.0-fast; 128K-token context; Matryoshka dimensions 256/512/768/1024/1536/2048; text, image and fused text+image input; 100+ languages
- Shared embedding space: index with Pro and query with Fast (or Pro) without re-indexing
- ViDoRe V3 average (Cohere): Embed 5 Pro 85.8, Fast 84.5; Voyage 4 Large 83.7, Gemini Embedding 2 83.2. Some secondary reports give Fast 84.7 and OpenAI text-embedding-3-large 75.5
- Fast delivers ~2.4x the throughput of Pro (~377 vs ~160 documents/s, per Cohere)
- Pricing: Pro $0.12 and Fast $0.08 per 1M text tokens; images $0.40 per 1M tokens (Cohere blog)
- Available on the Cohere API, Model Vault, Microsoft Foundry and Amazon SageMaker, and in Cohere North; private deployment via vLLM
What happened
Cohere launched its fifth-generation embedding family, which replaces Embed v4. The main design choice is two models that are interoperable: a high-quality Pro model for offline indexing and a cheaper, faster Fast model for live queries against the same vectors.
Why it matters
Embeddings power enterprise search and RAG. Benchmarks on visually rich documents (ViDoRe) are where Cohere, Voyage and Google now compete. A shared embedding space across tiers is an unusual, practical feature that lowers re-indexing costs.
Changelog
- 2026-10-01: created (leads run, 07:40 completion)
Models
- Cohere Embed 5 Fast Cohere · current
- Cohere Embed 5 Pro Cohere · current
Related events
Sources (3)
- officialCohere: Embed 5, frontier embedding models for enterprise
- docsCohere docs changelog: Announcing Cohere's Embed 5 models
- pressThe New Stack: Cohere's faster query model barely dents retrieval quality in its tests
id: 2026-09-30-cohere-embed-5 · updated 2026-10-01 · open in the interactive timeline