AI news6
Claude Duet opens a US research preview for proactive work
Anthropic opened the Claude Duet waitlist on 21 October for US Pro and Max subscribers. Built around Opus 5.5, the app checks connected accounts and works on its own cloud computer while users are away. It shares existing plan limits and requires approval before sending messages, making payments, signing documents or deleting data.
claude.com ↗Why it matters
This new proactive app extends beyond earlier Claude workspace features, with attention management and background usage becoming practical constraints.
Recursive releases a 40-billion-parameter model for local coding
Recursive’s Flint activates 2.5 billion of its 40 billion parameters per token and is available under Apache 2.0. Its 21 October release is corroborated by launch partner OpenHands. Recursive reports more than 125 tokens per second on an M4 Pro MacBook Pro at short contexts, using its optimised inference engine.
recursive.com ↗Why it matters
This turns automated research into a released local coding model, with performance claims developers can test on their own hardware.
Google Cloud adds three OpenAI models in private preview
Google Cloud’s 22 October announcement brings GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna to selected enterprise customers in private preview. Purchases through Google Cloud Marketplace can draw on existing cloud commitments. Codex workflows can connect for inference, while Google Antigravity and Gemini for business integration are still planned.
cloud.google.com ↗Why it matters
This new distribution agreement changes enterprise procurement options, beyond the previously covered model releases and Google infrastructure updates.
LiteLLM removes a credential-stealing release after a supply-chain attack
LiteLLM says its 21 October release pipeline compromise put malicious version 1.107.0 on PyPI from 10:39 pm AEDT to 12:23 am the next day. Affected Docker tags remained compromised until 1:20 am AEDT on 22 October. The maintainer reports credential theft and remote-code tools in those artifacts, while its private enterprise registry was unaffected.
docs.litellm.ai ↗Why it matters
The different package and container exposure windows make release provenance an operational concern for teams connecting many AI providers.
NVIDIA adds Agent Knowledge Foundation and Kumo models to its toolkit
NVIDIA’s 21 October expansion combines Agent Knowledge Foundation, new Kumo structured-data models and its existing agent runtime. Its supply-chain team is using the foundation with Palantir’s platform. NVIDIA also reports SAP’s fine-tuned reranker running nearly three times faster and ServiceNow cutting reranking latency by nearly a quarter.
blogs.nvidia.com ↗Why it matters
The new enterprise-data layer and reported deployments extend beyond the earlier NVIDIA model release, tying agent infrastructure to specific operational workloads.
Ant Group and vLLM open-source Humming for quantised inference
The 21 October Humming release supports dense and mixture-of-experts inference on NVIDIA GPUs from Turing onwards, plus compatible devices. In one uncached H200 compilation test, its authors report 1.82 seconds with Humming and NVCC, compared with 23.35 seconds for CUTLASS and NVCC. These are kernel-level measurements rather than production application savings.
vllm.ai ↗Why it matters
A shared kernel library can reduce the work needed to support different quantisation formats and hardware generations.
AI in the nonprofit sector4
RBC backs a Canadian programme for nonprofit-designed AI agents
The Human Feedback Foundation announced its Agentic AI Design Exchange on 21 October with RBC support. Canadian nonprofit cohorts are scheduled to start in autumn 2026, designing agents with fairness, accountability and human oversight. Selected designs are intended for a Creative Commons library, while the announcement gives no funding amount or application deadline.
globenewswire.com ↗Why it matters
The programme adds a specific capacity-building route for nonprofits and a planned resource other organisations can adapt.
AskTBP turns more than 100 philanthropy resources into a chatbot
The Trust-Based Philanthropy Project launched AskTBP on 21 October. It draws on more than 100 published articles, toolkits, webinar transcripts and case studies to answer questions about trust-based giving. The tool is intended mainly for funders, but nonprofit leaders and other donors can also use it.
trustbasedphilanthropy.org ↗Why it matters
A bounded knowledge base gives grantmakers a way to find practical guidance without asking grantees to provide all of that education.
ARIA’s cybersecurity call has a 1 November Sydney cut-off
ARIA’s £59 million Safeguarded AI programme accepts cybersecurity proposals from academia, industry and nonprofits. The next review cut-off is 1 November at 1 am AEDT, equivalent to 31 October at 14:00 GMT, for work using AI to develop machine-checked security proofs. Funded teams work in eight-week sprints. The funding page does not specify an individual award size.
aria.org.uk ↗Why it matters
This live call is relevant to specialist nonprofit security research teams, rather than general organisational AI adoption.
BUILD discussion calls for African influence over AI ownership
Voice of Nigeria’s 21 October report from BUILD 2026 in Abuja describes Luminate and the MacArthur Foundation calling for greater African influence over AI infrastructure and decisions. Melanie Hui urged philanthropy to support researchers, advocates, journalists and community organisers, while John Palfrey argued for broader participation in AI governance. The report announces no new grant programme.
von.gov.ng ↗Why it matters
This new event-specific policy discussion concerns ownership and public-interest technology, distinct from the previously covered Humanity AI funding call.