Morning Brief: October 25, 2026 — 3 AI developments worth knowing

Three developments from October 20–22: NVIDIA adds visual document retrieval, GitHub identifies custom Copilot tools in usage reports, and an opt-in dashboard compares adoption with peers.

View sources
In this article
  1. 1. Nemotron 3.5 makes document images searchable
  2. 2. Copilot usage reports identify custom tools by name
  3. 3. Copilot adds opt-in comparisons with peer organizations
  4. Sources & method

The morning view

NVIDIA expands what a search system can retrieve; GitHub expands what administrators can see about Copilot use. Client coverage and peer participation still limit the reporting.

1. Nemotron 3.5 makes document images searchable

Development:

NVIDIA’s Nemotron 3.5 Embed 8B adds visual document retrieval to its text-search predecessor. The model card dates release to October 20; NVIDIA explained the update on October 21. Developers can search pages containing charts, tables and diagrams using text queries.

Downloadable weights permit commercial and noncommercial use under OpenMDW 1.1. NVIDIA says existing text indexes can be reused because the model preserves its predecessor’s text embeddings. The documented example requires a CUDA GPU. Embedding is part of a wider workflow: NVIDIA’s NeMo Retriever Library also extracts document content and stores searchable vectors. NVIDIA’s accuracy comparisons remain vendor evidence.

Our read: Has teeth. The downloadable model gives search developers a concrete way to retrieve evidence from visually complex documents.

Read our full coverage: NVIDIA adds document-image search to Nemotron 3.5 retrieval models

Source 1Source 2Source 5Source 6Source 7

NVIDIA NeMo Retriever diagram showing document extraction, embeddings, vector storage and query retrieval; background for the lead Nemotron 3.5 Embed item.
Modified by AI Toolcast: resized and converted from PNG to WebP, without cropping. NVIDIA’s NeMo Retriever workflow illustrates background for this edition’s lead Nemotron 3.5 Embed item. It is not the new model’s architecture and does not illustrate the other items. Licensed under Apache 2.0; read or save the complete license using Source 7 in Sources & method. Source: NVIDIA. © Copyright 2023-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved..

2. Copilot usage reports identify custom tools by name

Development:

GitHub’s October 22 update names an organization’s skills, custom agents, MCP servers, slash commands and plugins in Copilot reports instead of grouping them under generic labels. MCP connects AI tools to external services. Each breakdown now includes up to 25 entries, up from five, with remaining activity grouped together.

Eligible owners, billing managers and custom metrics viewers need the usage-metrics policy enabled. Coverage depends on client and version: app, prompt-mode and Agent Client Protocol reporting require Copilot CLI 1.0.97 or later; Copilot SDK reporting requires 1.0.21 or later. Built-in editor chat telemetry does not report names. Blocked telemetry can undercount use, and older report days retain generic labels.

Our read: Has teeth. Administrators gain a more useful way to identify which internally developed tools are finding users.

Source 3

3. Copilot adds opt-in comparisons with peer organizations

Development:

GitHub added industry comparisons to its Copilot impact dashboard on October 22. Eligible enterprises and independent organizations can opt in to compare agent adoption and relative pull-request merge rates with participating peers. Owners select company size and industry; small industry groups can broaden to all industries of that size.

Participation is off by default and requires the usage-metrics policy. Enterprise-owned organizations participate through their enterprise; enterprises used solely to assign Copilot Business licenses are excluded. Comparisons refresh daily and stay in the dashboard, without API or export access. Data-residency customers are compared only with participants in the same region.

Our read: Watchlist. Rankings need enough participating peers and a calculable score; early adopters may see a broader comparison or no ranking.

Source 4

Sources & method

AI-assisted morning roundup based on primary announcements, model documentation and release records. Selection and assessments are editorial judgments about documented capabilities and access, not popularity rankings or hands-on tests.

  1. NVIDIA: Nemotron 3.5 Embed, Rerank, and Parse release announcementPublished October 21, 2026
  2. NVIDIA: Nemotron 3.5 Embed 8B BF16 model cardAccessed October 25, 2026
  3. GitHub: Custom skills, agents, and MCP servers in usage metrics APIPublished October 22, 2026
  4. GitHub: Industry comparisons in the Copilot impact dashboardPublished October 22, 2026
  5. NVIDIA: NeMo Retriever Library overviewAccessed October 25, 2026
  6. NVIDIA: NeMo Retriever documentation licensing and third-party noticesAccessed October 25, 2026
  7. Full Apache License 2.0 text for the NVIDIA NeMo Retriever diagram — read or save a copyAccessed October 25, 2026

Source pages reviewed October 25, 2026. Analysis and suggested evaluation methods are AI Toolcast’s own.

This AI-assisted, source-reported roundup contains no sponsored placement or affiliate links. Products were not tested for this edition; vendor performance claims have not been independently reproduced.

Found a possible error? How to report a correction.