How this page is kept. Every morning a script checks Hugging Face and each vendor’s model, pricing and deprecation pages for changes. A person reads the vendor’s own page before anything is written here, and figures are re-checked against it daily. Follow it by RSS.
The library most people use to run open rerankers shipped 6.3.0. The release is mostly about multi-vector (ColBERT-style) training; nothing changes for scoring with CrossEncoder. We ran this site’s self-hosting quick start, the GTE and mxbai examples and the ColBERT snippet on it, unchanged.
Source: release notes. On this site: self-hosting guide.
5 Nov 2026 — Cohere prices Model Vault by the GPU
Cohere’s pricing page now lists Model Vault, its dedicated deployment, as an hourly rate per GPU type: from $3.00 per GPU-hour for an L4 on demand, times the number of replicas. The per-model rates it listed for Rerank 5 in October are gone. Pay-as-you-go Rerank 5 prices are unchanged.
Source: Cohere pricing. On this site: Cohere Rerank pricing.
3 Nov 2026 — Pinecone deprecates pinecone-rerank-v0
Pinecone’s docs now mark its own reranker as deprecated: it still serves requests, but shouldn’t be used for new projects. No shutdown date is given. The two other rerankers Pinecone hosts, bge-reranker-v2-m3 and cohere-rerank-4-fast, are unaffected.
Source: Pinecone model page. On this site: reranking inside Pinecone.
30 Oct 2026 — Pinecone halves a limit on cohere-rerank-4-fast
cohere-rerank-4-fast on Pinecone now takes 4,096 tokens per document, half of what Pinecone’s docs listed before. Longer documents are truncated, so check your chunk sizes if you rerank long passages there.
Source: Pinecone model page. On this site: reranking inside Pinecone.
29 Oct 2026 — Nemotron-3.5-Rerank-8B becomes a NIM
NVIDIA’s NIM release 2.4 serves the 8B reranker as the nemotron-rerank container. It is capped at 10,240 tokens per query-document pair by default, and the limit can be raised to the model’s full 32,768. The older llama-nemotron rerankers stay on NIM 2.3.
On this site: NVIDIA Nemotron rerankers.
28 Oct 2026 — Hindsight releases hs-rerank-1
A small multilingual cross-encoder under the MIT licence: 114M parameters, of which 98M are the embedding table, 8,192 tokens of context, eight languages including Chinese, Japanese and Korean, and an int8 ONNX file for CPUs. Its model card shows results only as an image, so there is no published number to compare it by yet.
Source: model card. On this site: models table.
21 Oct 2026 — Cohere releases Rerank 5
Two models, rerank-v5.0-pro (about 2B parameters) and rerank-v5.0-fast (about 500M), both with 32,768 tokens of context. New: calibrated 0–1 scores, so one threshold works across queries, and Context Focus, which returns the passages inside a document that answer the query. Pricing moved from per search to per token: $0.05 per million tokens for Pro and $0.02 for Fast. Rerank 4 is still served, but its prices left the pricing page.
Source: Cohere’s announcement. On this site: Cohere Rerank review.
20 Oct 2026 — NVIDIA releases Nemotron-3.5-Rerank-8B
An 8.38B reranker that scores text, page images or both, with 32,768 tokens of context and a licence that allows commercial use. In NVIDIA’s own run it scores 67.22 on ViDoRe V3 with page images, ahead of Qwen3-VL-Reranker-8B; other labs report different figures for the Qwen model, and our page sets them side by side.
Source: model card. On this site: NVIDIA Nemotron rerankers.
30 Sep 2026 — Voyage releases rerank-3
A drop-in upgrade from rerank-2.5 at the same price, $0.05 per million tokens, with 32,000 tokens of context. It still takes natural-language instructions. Voyage reports its largest gain on long documents, 3.35% over rerank-2.5 in its own evaluation.
Source: Voyage’s announcement. On this site: Voyage Rerank review.