What changed in AI, and what it means for your work.

NVIDIA has released open, commercially available Nemotron 3.5 retrieval models on Hugging Face: Embed in 8B and 1B sizes for searching document text and images such as tables and charts, an 8B Rerank model that reorders search results, and Parse, a 631M-parameter model that turns document images into structured text. NVIDIA says adding the reranker cut an agent's tokens per query by about 14% with the 8B embedder while improving retrieval accuracy, and that Embed 8B and Rerank each ranked first on ViDoRe v3.1 leaderboards as of 20 October 2026. Text retrieval was evaluated across 34 languages. Partners including Vectorize, Perplexity and Turbopuffer report strong results in their own tests, as quoted by NVIDIA.

Why it matters: If you build search or chat over PDFs, invoices or reports full of tables and charts, try Parse and Embed on a sample of your own documents and compare the results with your current setup.

NVIDIA, IIT Madras and its AI4Bharat lab have announced Nemotron-pretraining-Indic-v1, about 2.8 trillion tokens of training data across the 22 languages listed in India's Constitution, which they say will be released soon along with the open, GPU-accelerated pipeline used to make it. The data is built by translating curated English material, such as maths, code and question-answer content, into Indian languages, with native script for all 22 languages and romanised text for twelve, and each output checked against its source. Adding 10 billion Indic tokens from a twelve-language subset plus natural Indic text to Nemotron 3.5 Super VL improved its Indic benchmark scores by 4.98 to 9.82 percentage points, though the authors note this does not isolate the synthetic data's effect. Sorted take: if it ships as described, this is a large open boost for anyone building Indian-language models, but translated data still lacks local knowledge.

Why it matters: If you build chatbots, support tools or search for customers who use Indian languages, this corpus and recipe could make fine-tuning in those languages much cheaper. Watch for the release and test it on your own language mix.

At the Open Compute Project Global Summit in San Jose, NVIDIA said on 12 October 2026 that it has contributed key Vera Rubin rack and compute tray designs to the Open Compute Project through its open MGX architecture. NVIDIA says the updated MGX 1.2 specification adds blind-mate liquid cooling and a liquid-cooled 5,000A busbar. It also published a Vera CPU rack design aimed at agent workloads such as tool calling, orchestration and sandboxed execution. Alongside it, NVIDIA, AMD, Google and Meta published the Open Data Center Specification v0.7, a shared blueprint for AI data center power, cooling and layout. Google, Meta, Microsoft and NVIDIA published an OCP Grid Resilience Specification, which sets out how large AI sites should stay connected during grid disturbances and restore to near-full load within less than two seconds. NVIDIA also released open-source software for running clusters, and Cisco announced plans with NVIDIA for USFP, a new optical module form factor that the companies say will deliver four times the bandwidth density of 200Gb/s-per-lane pluggable optics. The performance claims are the companies' own.

Why it matters: If you plan or rent AI data center capacity, open rack and facility specs backed by several hyperscalers should make it easier to build and buy compatible equipment instead of waiting for one vendor's designs.