AI models
21 Oct ~06:10 – ~18:15 CEST · Lab blogs, Hugging Face and lab X accounts. The morning pull was quiet. Both releases that cleared the bar today are retrieval models, the plumbing that decides what an agent gets to read. The Nemotron 3.5 Embed/Rerank repos had been quietly public on Hugging Face since 20 Oct, and Cohere's Rerank 5 docs surfaced late on 20 Oct, but both official launches came today. Claude Fable 5.5, Kimi K3.1 and Gemini 4 Carbon are still leaks and rumours, and MiniMax's M3.1 Flash (covered 19 Oct) still has its weights due on 23 Oct.
NVIDIA — Nemotron 3.5 Embed, Rerank and Parse open retrieval models (GTC Berlin) (21 Oct, ~12:00 CEST)
At its GTC Berlin keynote, NVIDIA formally launched a new generation of Nemotron Retriever models: Nemotron 3.5 Embed in 8B and 1B sizes (BF16, plus an NVFP4 build of the 1B), Nemotron 3.5 Rerank 8B, and Nemotron 3.5 Parse, a 631M-parameter document-extraction model. The embedders are multimodal. They index a document page as text, as an image, or both, so tables, charts, infographics and schematics become searchable alongside the prose. Text retrieval was evaluated across 34 languages and image retrieval across 18. The models are built on Mistral's Ministral 3 backbones with a frozen Nemotron 3 Embed language tower, so existing Nemotron 3 text indexes don't need re-embedding. The reranker is an 8B cross-encoder with 32K context that scores query–document pairs across text and images.
NVIDIA says Embed 8B tops the ViDoRe v3.1 leaderboard for dense retrievers (61.17 nDCG@10). The distilled 1B keeps about 95% of that score, and the NVFP4 build runs about 27% faster on GB200 with practically no accuracy loss. Rerank 8B leads a new ViDoRe v3.1 reranking leaderboard, launched with the MTEB team, narrowly ahead of Qwen3-VL-Reranker-8B. In NVIDIA's agent tests, adding the reranker cut agent token use per task by about 13–14%. Parse ranks first on ParseBench among sub-1B models at about 10–16 pages per second on one H100. Weights are open for commercial use under the OpenMDW-1.1 licence on Hugging Face, with NIM microservices on build.nvidia.com and hosting via Baseten, DeepInfra, FriendliAI and OpenRouter. All benchmarks are NVIDIA's own, plus partner quotes from Perplexity, Turbopuffer and Vectorize. NVIDIA's open-models post also named Kumo models for tabular and time-series prediction, the RFD4-Proteina protein-design model with the Baker Lab, new Cosmos tooling, Alpamayo 2 Edge (weights later this month) and Nemotron 3.5 Bolt E4B (due Friday 23 Oct).
Why it matters: as agents run more searches, retrieval quality now sets a big share of both answer quality and token bills. NVIDIA is giving away a full open, multilingual, image-aware retrieval stack, which pressures paid embedding and reranking APIs.
Cohere — Rerank 5 Pro and Fast (21 Oct, ~17:00 CEST)
A few hours after NVIDIA, Cohere released Rerank 5, its new reranking family, in two tiers. Rerank 5 Pro (about 2B parameters) aims for top quality on reasoning-heavy queries and long, domain-specific documents. Rerank 5 Fast (about 500M) targets high-volume search and agents that fire off many queries. Cohere says Fast already beats its previous flagship, Rerank 4 Pro. Both handle 100+ languages, a 32K-token window and semi-structured data such as JSON and tables. They add two practical features: Context Focus, which returns the most relevant passages inside each document as scored spans in the same call so less text has to go to the LLM, and calibrated 0–1 relevance scores, so a single cut-off works across queries and domains.
On Cohere's numbers, Pro ranked first on 39 of 46 composite datasets across ViDoRe V3, BEIR, BRIGHT and nanoMTEB. On the reasoning-heavy BRIGHT benchmark it scores 82.2, 5.8 points ahead of Voyage AI Rerank 3 and 12.4 ahead of Rerank 4 Pro. It also leads Voyage in all six priority business languages Cohere showed. These scores use Cohere's own RCP-nDCG@10 variant, so they aren't directly comparable with NVIDIA's figures. The models are closed weights. API pricing is $0.05 per 1M tokens for Pro and $0.02 for Fast, and they're also available via Model Vault, Microsoft Foundry, Amazon SageMaker, North and private VPC/on-prem deployments. Upgrading needs no re-indexing. Rerank 5 completes Cohere's fifth-generation search stack alongside Embed 5 and Parse 5.
Why it matters: two big retrieval launches on the same day show where the enterprise fight has moved. The question is less which model sounds smartest and more which one finds the right evidence for the fewest tokens.