AI news5
Nanbeige releases a 33B agent model with a 512K context window
Nanbeige’s new 4.3-33B model is available under Apache 2.0, with support for up to 524,288 context tokens. The team positions it as a teacher for future smaller agents, including a planned 3B model.
huggingface.co ↗Why it matters
Open weights give teams another option for running coding and work agents on their own infrastructure, while the smaller model remains a future release.
Apple announces A20 Pro iPad mini, arriving 6 November
Apple’s 26 October announcement introduces an 8.5-inch OLED iPad mini with the A20 Pro chip. Preorders start 28 October, with availability from 6 November.
apple.com ↗Why it matters
This separate launch from the earlier HomeView announcement expands Apple’s hardware for on-device AI; Siri AI still has beta, language and regional limits.
Meta launches US$35 billion in equity offerings for its AI push
Meta announced roughly US$35 billion in concurrent underwritten offerings on 26 October. It also plans a separate US$60 billion at-the-market programme, whose timing and terms remain undetermined.
investor.atmeta.com ↗Why it matters
Unlike the earlier Sundial acquisition, this financing plan shows the scale of capital sought for AI infrastructure; it is not US$95 billion already raised.
Atos launches Genesis Intelligence and its Codex operations platform
Atos announced Genesis Intelligence on 26 October, with deployment options including sovereign cloud, on-premises and air-gapped environments. Its Atos Codex platform connects operational data and AI reasoning to actions under controls defined by each organisation.
atos.ai ↗Why it matters
The launch targets regulated organisations that need control over where AI runs and which actions require human oversight; customer outcomes remain unverified.
Oracle announces a four-layer blueprint for enterprise AI controls
Oracle’s 26 October Deep Trust AI Architecture puts safeguards across software, data, data access and application development. The blueprint combines database access rules, controlled data APIs and the APEX application runtime.
oracle.com ↗Why it matters
This new architecture, separate from earlier Oracle financing and education stories, aims to enforce permissions beneath individual agents rather than rely on each generated application.
AI in the nonprofit sector4
Claude adds a nonprofit plugin built with about 30 organisations
Anthropic released the Claude for Nonprofits plugin on 26 October, developed with about 30 nonprofits. It supports fundraising and operations and installs automatically at no additional cost for verified nonprofits on Team and Enterprise plans.
claude.com ↗Why it matters
This is a new workflow package beyond the previously covered Claude Corps programme; the proposed Taproot extension remains a plan.
BDO survey: 70% of nonprofit respondents have formally adopted AI
BDO’s 26 October report surveys 250 nonprofit leaders, with 70% reporting formal AI adoption in at least one area. It says 93% use AI in some capacity and 58% manage oversight within IT.
bdo.com ↗Why it matters
The gap between informal use and formal adoption gives nonprofit boards a concrete reason to review accountability; these are survey responses, not a sector-wide census.
Google.org awards Inria Chile US$2.5 million for ocean AI
Inria Chile’s 14 October announcement confirms US$2.5 million from Google.org for MOFOMO over 36 months. The project will combine physical, biological and genomic observations in an open ocean model intended to support climate forecasts.
mofomo.inria.cl ↗Why it matters
This specific award advances beyond the previously covered AI for Science funding call, while reliable ocean predictions remain a research goal.
Google.org commits US$15 million to Samriddhi’s MSME AI programme
The programme operator, the^delta, announced Google.org’s US$15 million commitment on 14 October. Samriddhi aims to support 100,000 Indian MSMEs over three years; four AI Grand Challenges are planned to launch in January 2027.
linkedin.com ↗Why it matters
This distinct livelihoods programme expands beyond earlier Google.org awards, with future application rules and outcomes still to be established.