Model Card: Cohere Embed 5

🌏 中文版

Model Details

Key Capabilities

• On ViDoRe V3 (retrieval over visually rich enterprise documents — financial filings, technical manuals, regulatory material, government reports), Embed 5 Pro averages 85.8, up 8.8 points from Embed 4's 77.0, with the biggest gains in HR (+11.4) and industrial (+10.3) domains
• First model family evaluated with RCP-nDCG@10, a new Cohere methodology that scores retrieved documents against query-specific relevance criteria instead of a fixed label set, which Cohere says better reflects quality on a team's own corpus
• Pro and Fast share a single embedding space: you can index with Pro and query with Fast (or vice versa). Across 40 development datasets, cross-model combinations lose only 1.6–2.7% on average versus same-model baselines, with no index rebuild required
• Supports Matryoshka embeddings plus float/int8/binary formats: a 2048-dimension float32 vector (8KB) can shrink to a 256-dimension binary vector (32 bytes) — a 256x reduction. Across 100 million chunks, that cuts vector storage from roughly 819GB to 3.2GB; Cohere recommends 1024-dimension int8 vectors as the typical quality/cost sweet spot

Benchmark Results

⚠️ All figures above are Cohere's own benchmarks; there's no independent third-party reproduction yet. RCP-nDCG@10 is a Cohere-designed methodology introduced with Embed 5 — Cohere published the annotations and code for outside review, but the evaluation set itself was still designed and curated by Cohere.

vs. Previous Generation / Competitors

Compared with Embed 4, Embed 5's biggest gains aren't in plain English text retrieval — they're in documents where structure itself carries meaning: tables, charts, multi-column layouts, scanned financial pages that older "convert to text, then embed" pipelines tend to mangle. The 8.8-point average gain on ViDoRe V3, with HR and industrial domains both up more than 10 points, shows this release is mainly about closing the gap on non-plain-text document retrieval.

Against competitors, Embed 5 Pro leads on three vertical fronts — visually rich documents, financial documents, and multilingual retrieval (especially Middle Eastern and South Asian languages like Farsi, Telugu, and Hindi) — typically by 1–3 points over Voyage 4 Large and Gemini Embedding 2. The margin over OpenAI's text-embedding-3-large is much larger (21.4 points on FinanceBench). Notably, the announcement never cites a general-purpose benchmark like MTEB, which suggests Cohere is positioning this release around enterprise vertical differentiation rather than claiming a blanket new state of the art on general retrieval.

On pricing, Embed 5 splits indexing and querying into two price points — Pro costs 50% more than Fast, but Fast delivers 2.4x the average document throughput. That's different from most embedding models, which charge one price regardless of use case; Embed 5 effectively bakes the two-stage nature of a retrieval pipeline (batch indexing, where cost sensitivity is lower and quality matters most, vs. live querying, where latency and cost matter most) directly into the pricing.

What This Means for Agent Development

Embed 5 isn't a generative model, but the quality ceiling for most RAG or agentic retrieval pipelines is often set by the retrieval layer, not the generation layer — if what gets retrieved is irrelevant or missing structure, no amount of LLM capability can fix that downstream.

• If you're building an agent loop that retrieves repeatedly (multi-turn search, a deep-research agent that narrows scope step by step): the shared embedding space fits directly — index once with Pro to get quality right, then use the cheaper Fast for every subsequent query round, without sacrificing index quality or rebuilding the index just to switch models
• If you're building a knowledge-base agent over financial reports, technical manuals, or regulatory documents where the information lives in tables, charts, or layout: Embed 5 Pro supports page-image and fused text-image embeddings directly, letting you skip the "OCR or convert to text first" step and avoid the structural loss that step tends to introduce
• If storage and infrastructure cost matter to you: int8/binary quantized embeddings combined with Matryoshka can cut vector storage to 1/256 of the baseline at scale (hundreds of millions of chunks) — a more concrete win for teams running their own vector databases than the benchmark gains themselves
• Not a fit: lightweight, plain-English FAQ retrieval — Cohere didn't bring out a general benchmark like MTEB for this release, so if your use case is simple text-only semantic search, a cheaper existing open-source embedding model may already be good enough without paying for Embed 5's vertical-domain advantages

Today's Takeaway

Embedding model releases often get waved off as "minor infrastructure updates" because they can't demo a conversation the way a chat model can. But Embed 5 turns "index with the expensive Pro, query with the cheap Fast" into the core of its product design — which tells you the vendor itself recognizes that a retrieval pipeline's cost structure is fundamentally different from a generative model's: indexing is a one-time batch cost, while querying is a cost that accumulates with every loop of an agent's retrieval calls. Those two stages deserve different cost/quality tradeoffs by design, not a single model trying to cover both.

References

• Cohere Blog: Embed 5 — Frontier Embedding Models for Enterprise
• Cohere Docs: Cohere's Embed Models (embed-v5.0-pro / embed-v5.0-fast spec table)
• tao.media: Cohere Launches Embed 5 Pro and Fast Frontier Embedding Models
• digitalapplied: Cohere Embed 5: Pro and Fast Embedding Models Compared