AI news6
Salesforce unveils Koa, a CRM reasoning model built with Nvidia's Nemotron
At Dreamforce 2026 in San Francisco, Salesforce introduced Koa, its first reasoning model purpose-built for CRM and the Agentforce platform, developed with Nvidia and based on Nvidia's Nemotron 3 Super open model. Salesforce post-trained it on a proprietary synthetic dataset modelled on knowledge drawn from nearly three decades of CRM deployments. Koa is available now to selected Agentforce pilot customers, with general availability expected in US regions in winter 2026, and the two companies are separately integrating Nemotron-based models into Missionforce, Salesforce's platform for government and regulated organisations, including air-gapped deployments.
newsbytes.ph ↗Why it matters
The first major SaaS vendor to ship a domain-specific reasoning model on open weights signals that enterprise AI differentiation is shifting from foundation models to proprietary workflow data and process, not raw capability.
OpenAI confirms weeks of joint AI safety work with Anthropic and Google DeepMind
OpenAI's global policy chief Chris Lehane said the company has been coordinating on AI safety with Anthropic and Google DeepMind for several weeks, and argued the three labs do not need an antitrust waiver to do so. The disclosure follows an essay by Anthropic CEO Dario Amodei calling on the industry to slow frontier development, which drew public support from OpenAI's Sam Altman, Google's Demis Hassabis and Elon Musk. OpenAI is backing the FRONTIER Act, which would require frontier labs to submit models for audit by independent evaluators, while a separate bill from Senate Majority Leader John Thune with Senators Ted Cruz and Amy Klobuchar would give the Commerce Secretary pre-release audit powers.
bloomberg.com ↗Why it matters
After years of competition-first positioning, the three largest US labs are now openly coordinating on safety standards, which moves independent evaluation from a research talking point to live legislation.
Arcee AI hits a $1B+ valuation to build American open-weight frontier models
Arcee AI raised a Series B led by Vista Equity Partners, Cambium Capital and Emergence Capital, with Hitachi, Microsoft's M12, Wipro and others participating, valuing the San Francisco company at more than $1 billion. Arcee built its 400-billion-parameter Trinity Large mixture-of-experts model, plus its entire 2025 lineup, for roughly $20 million, and says Trinity Large was the first permissively licensed frontier-class model developed end to end in the United States since Meta stopped releasing Llama. It will use the funding to train the next Trinity generation, expand work with the US Department of Energy and its 17 national laboratories on the open Genesis-Science-1 scientific model, and build deployment tooling for organisations running open models in production.
globenewswire.com ↗Why it matters
Capital-efficient, openly licensed frontier models built outside the hyperscalers give smaller organisations a credible alternative to closed APIs, which matters most to cost-sensitive public and nonprofit users.
TypeSafe AI exits stealth with $40M and Jev, a frontier model built for software rather than chat
TypeSafe AI, founded in 2024 by former OpenAI researcher Diogo Almeida, emerged from stealth with $40 million in seed funding led by DCVC and a new model called Jev. Instead of returning conversational text, Jev returns typed probabilistic decisions such as yes-or-no probabilities or scores on a defined scale, each with a confidence measure so developers can set thresholds for when software acts automatically, seeks more information, or defers. The company says Jev can respond in under 100 milliseconds and up to 100 times faster and cheaper than comparable frontier models, and is offering it now through an early-access waitlist.
techstartups.com ↗Why it matters
If machine-to-machine models become a distinct class from human-facing chatbots, the cost of automating judgment-heavy back-office work drops sharply, and that is exactly where nonprofits carry the most manual load.
Mistral and Mozilla bring open, private AI to Firefox browsing
Mozilla's AI browsing assistant, Firefox Smart Window (beta), is now powered by Mistral models, the two companies announced. Smart Window helps users work through complex searches, recall pages they clicked away from, and source information across open tabs, and the partners state conversations are not stored by default under a no-data-retention agreement. The rollout covers users in France and North America first, with the United Kingdom and Germany expected later this year, and Mistral says its models are tuned for regional languages, dialects and cultural context.
mistral.ai ↗Why it matters
A privacy-first, multilingual assistant shipped inside a mainstream browser puts open-weight models directly in front of ordinary users, not just developers, which is the distribution problem open models have struggled with.
LawZero receives up to CAD $300M from Canada and Germany for safe-by-design AI
Canada and Germany committed up to CAD 150 million and EUR 100 million respectively to LawZero, the not-for-profit AI safety lab founded by Turing Award winner Yoshua Bengio, announced at Montreal's All In AI conference. The funding will grow its research team, secure sovereign compute in Canada with Hypertec and 5C, and open a Berlin office, backing a "Scientist AI" approach designed to reason transparently and produce evidence-based outputs without pursuing goals of its own.
newswire.ca ↗Why it matters
Two governments are funding a non-profit alternative to the frontier labs on the explicit thesis that safety can be designed in from the start, which gives the public-interest AI sector a well-resourced reference point.
AI in the nonprofit sector5
Humanity AI opens its first $10M grant call, US nonprofits have until 21 October
Humanity AI, the five-year $500 million philanthropic coalition of ten US foundations, opened its first open call on 10 September, directing $10 million to community-led projects shaping how AI is used in education, labour and the economy, and arts and culture. US nonprofits can apply for grants of $75,000 to $1 million, with applications due 21 October 2026, 4:59 pm Pacific, and awards expected in spring 2027. The call is administered by Lever for Change, and the coalition includes the Ford, MacArthur, Mellon, Mozilla, Packard, Doris Duke, Lumina, Kapor and Siegel foundations plus Omidyar Network.
techrepublic.com ↗Why it matters
Most of these funders do not accept unsolicited proposals, so a pooled open call is one of the few direct routes for small and mid-sized nonprofits into serious AI funding this cycle.
Gates Foundation becomes the first philanthropic member of the Agentic AI Foundation
The Linux Foundation's Agentic AI Foundation, the neutral home for the agentic AI stack including Anthropic's Model Context Protocol, announced the Gates Foundation as its first philanthropic member, with senior AI strategy adviser Matt Vasey taking a board seat. The foundation will bring global health, education, agriculture and economic opportunity perspectives into the standards work, and has established a fund to support African developers attending MCP Dev Summit Nairobi.
linuxfoundation.org ↗Why it matters
The standards body shaping agentic AI now has a philanthropy at the table, which is the earliest point at which low-income-country and civil-society needs can influence the stack rather than retrofit onto it.
OpenAI Foundation launches Public Data for Health with a $40M cancer-vaccine data grant
The OpenAI Foundation announced Public Data for Health, an effort to fund the creation of the high-quality scientific datasets that current AI models lack. Its first round includes $40 million to the University of North Carolina at Chapel Hill to collect data on novel cancer vaccines, a grant to OpenAdmet for competitions in which researchers try to predict drug effects, and $500,000 to 1Day Sooner for a biotech archive of clinical-trial volunteer data. The foundation gave $100 million to the Common Health Coalition in August and says it aims to distribute $1 billion by the end of the year.
technologyreview.com ↗Why it matters
Funding the creation of datasets rather than access to models attacks a bottleneck small research nonprofits cannot solve alone, and it is a repeatable template for how AI funders can buy public goods.
German investor Jan Beckers commits EUR 50M to a new AI Safety Fund for nonprofits
Jan Beckers and the Global Technology Risk Foundation launched the GTR AI Safety Fund, a EUR 50 million multi-year program for individuals and non-profit organisations working on the safe development of AI, with applications and nominations open on a rolling basis from 14 September. Support is tailored to each recipient and ranges from early research or compute funding to multi-year organisational backing, and the foundation has previously backed Apollo Research, the KIRA Center and the AI Whistleblower Initiative. For-profit companies making substantial contributions to risk reduction are supported separately through Beckers' family office, Ocean Investment.
live.deutsche-boerse.com ↗Why it matters
A rolling, low-bureaucracy funding stream for AI safety work in Europe answers the common complaint that safety talent has ideas and no early capital.
India's Rao Family Trust opens a Tech for Good grant for deployed AI, deadline 30 September
The Rao Family Trust (RAFT) opened a Tech for Good call for proposals for organisations using AI at the core of solutions to social challenges, offering up to INR 50 lakh and potential multi-year support, with a preference for work in Maharashtra across agriculture, healthcare, disability inclusion and frontline support. RAFT states that only organisations that have already deployed AI solutions with measurable impact will be considered, and that ideas, prototypes and pilots will not proceed. The deadline is 30 September 2026.
ngoportal.org ↗Why it matters
A funder requiring deployed, measured AI impact rather than a proposal is a useful signal for what "ready" now looks like in a developing-market nonprofit context.