Cohere's Embed 5 indexes and queries in one space

Cohere released Embed 5, a family of frontier embedding models in two tiers, Pro and Fast, that share a single embedding space. The practical consequence is that you can index with Pro and query with Fast without re-embedding the corpus, and Cohere reports the quality cost of that swap as small: Fast queries against a Pro index scored 98.4 normalized against a 100 baseline, and Fast on both sides scored 96.6.

Pricing is $0.12 per million text tokens for Pro and $0.08 for Fast, with image input at $0.40 per million tokens on both. Both tiers carry 128K token context, accept text, image and fused text-image input, cover more than 100 languages, and offer output dimensions of 2048, 1536, 1024, 768, 512 and 256 in float, int8 and binary formats with Matryoshka support. They are generally available on the Cohere API and Model Vault, on Microsoft Foundry and Amazon SageMaker, and can be self-hosted through vLLM.

The numbers worth comparing

On ViDoRe V3, Pro scores 85.8 and Fast 84.5, against Voyage 4 Large at 83.7, Gemini Embedding 2 at 83.2, Cohere’s own Embed 4 at 77.0 and OpenAI’s text-embedding-3-large at 75.5. Pro’s largest domain gains over Embed 4 are in human resources at +11.4 and industrial at +10.3.

On finance, Pro leads FinanceBench at 80.1, FinQA at 90.0 and ViDoRe V3 Finance at 85.0, averaging 3.3 points above Gemini Embedding 2 and 21.4 points above OpenAI’s text-embedding-3-large on FinanceBench. On parsed PDFs, Pro averages 84.8 against 83.6 for Voyage 4 Large and 80.8 for Gemini Embedding 2. On fused text-image retrieval, Pro scores 82.3 against Gemini Embedding 2’s 61.3. Fast runs at 377.3 documents per second against Pro’s 159.7, roughly 2.4 times faster.

The storage argument is where the smaller dimensions earn their keep: 2048-dimension float32 is 8 KB per vector, 1024-dimension int8 is 1 KB, and 256-dimension binary is 32 bytes, a 256 times reduction. Across 100 million chunks that is roughly 819 GB down to 3.2 GB. Cohere recommends 1024-dimension int8 as the middle point.

Two caveats on reading these tables. The RCP-nDCG@10 metric Cohere uses requires a two stage retrieval setup, so the scores reflect reranking quality rather than first stage retrieval performance, and Cohere flags that its upcoming fifth generation models may not always look strongest under legacy nDCG alone. Benchmark documents were parsed with Gemini 1.5 Flash, and the High Finance set is internal and Cohere annotated. Cross-model compatibility only works if both sides use the same output dimension, and the binary format trades some accuracy for the size win.

(Original source: Cohere)