AI news5
Mistral Medium 3.5 ships as 128B dense model with 256K context, agentic Le Chat Work mode and async cloud coding agents in Vibe
Mistral released Medium 3.5 on 29 April as an open-weights 128-billion-parameter dense model under modified MIT license, folding instruction-following, reasoning and coding into one architecture with a 256,000-token context window. The model scores 91.4% on τ³-Telecom and 77.6% on SWE-Bench Verified, and powers a new agentic Work mode in Le Chat that runs multi-step tasks with several tools at once. The same release made Vibe coding agents practical to ship in the cloud, with sessions spawnable from CLI or Le Chat, asynchronous execution in isolated sandboxes and live teleport from local sessions.
mistral.ai ↗Why it matters
A 128B open-weight model that matches frontier coding-agent benchmarks at meaningfully lower cost gives any nonprofit with technical staff a sovereign deployment option that needs no procurement conversation with the major US labs.
Salesforce Agentforce Operations enters general availability with back-office AI agents that complete work, not just orchestrate it
Salesforce announced Agentforce Operations general availability on 29 April, introducing AI agents that execute back-office tasks end-to-end, adapt as the business changes and only loop in humans when needed. The company also reported $800 million in Agentforce annual recurring revenue at fiscal-2026 close, up 169% year-on-year. New platform additions include Agentforce Builder for faster agent development, Agent Script for behaviour control, Agentforce Voice for natural conversation and Intelligent Context for grounding agents in unstructured data.
salesforce.com ↗Why it matters
Salesforce is now the largest pure-play agentic-AI revenue line in enterprise software, and any nonprofit running Salesforce-based fundraising, programme delivery or member-services back office should expect Agentforce-built workflows to become the default upgrade path through 2026.
Adobe Firefly AI Assistant launches in public beta with agentic orchestration across Photoshop, Premiere, Lightroom, Illustrator and Express
Adobe shipped its Firefly AI Assistant in public beta on 27 April for Creative Cloud Pro, Firefly Pro, Pro Plus and Premium plans. The assistant takes natural-language descriptions of a desired creative outcome and orchestrates multi-step workflows across Adobe's apps, drawing on more than 60 pro-grade tools including Auto Tone, Generative Fill, Remove Background and Vectorize. Adobe confirmed it will expand the assistant to use third-party AI models, with Anthropic's Claude as the first integration in development.
blog.adobe.com ↗Why it matters
Communications, marketing and storytelling teams at nonprofits running Adobe Creative Cloud now have a native agent capable of producing finished assets from a brief, and the Claude integration line means any AIH-aligned org running both stacks can keep brand-voice control while accelerating output.
NIST CAISI publishes formal evaluation of DeepSeek V4 Pro showing capabilities lag frontier by 8 months and persistent jailbreak weakness
The National Institute of Standards and Technology's Center for AI Standards and Innovation (CAISI) published its evaluation of DeepSeek V4 Pro in May 2026, finding the open-weights Chinese model trails frontier labs by approximately 8 months on capability benchmarks. The previous CAISI evaluation last September found DeepSeek complied with public jailbreak prompts in 95-100% of tests for phishing, malware and other restricted uses, versus 5-12% for comparable US models, and the V4 Pro evaluation extends that censorship-and-safety risk profile. CAISI flagged the V4 Pro's behaviour around CCP-aligned narratives as a procurement concern for application developers and US national security users.
nist.gov ↗Why it matters
This is the first formal US-government safety dossier on a 1.6T-parameter open-weights frontier-tier model that nonprofits might be tempted to self-host for cost reasons, and any board reviewing a "we run DeepSeek instead of Claude" proposal now has an authoritative document to consult.
Fortune confirms Big Tech 2026 AI capex tracking $700 billion across hyperscalers as no end to the buildout is in sight
Fortune reported on 30 April that combined 2026 AI infrastructure capex from the major hyperscalers is tracking $700 billion, sitting well above the $130 billion Q1 figure that defined last week's earnings season. Microsoft, Google, Amazon and Meta are now in a binding capacity-constrained phase, with Google explicitly flagging compute as the limiting factor on its data-centre buildout and Meta raising 2026 capex guidance to $145 billion. The piece argues there is no clear endpoint to the buildout cycle while frontier model training remains the gating workload.
fortune.com ↗Why it matters
A nonprofit budgeting AI tooling for 2027 needs to model both vendor pricing pressure (cost recovery on $700B) and supply allocation risk, because hyperscalers are now openly choosing which customer workloads to prioritise on constrained capacity.
AI in the nonprofit sector4
Salesforce Agentforce Nonprofit launches generally available with purpose-built AI agents for fundraising, programme delivery, volunteer ops and donor support
Salesforce announced Agentforce Nonprofit's general availability with purpose-built AI agents covering Prospect Research and Participant Management as GA today, Volunteer Capacity & Coverage in beta with GA early 2026, and Donor Support in spring beta with summer GA. The company paired the launch with Salesforce Accelerator Agents for Impact, an initiative providing technology, funding and expertise to help nonprofits build and customise their own AI agents. The Participant Management agent generates automated summaries of participant history, framed as cutting hundreds of administrative hours that staff can redirect to community work.
salesforce.com ↗Why it matters
This is the first time the dominant nonprofit CRM has shipped same-day-GA agents specifically for nonprofit core processes, and any AIH member running Salesforce NPSP now has a native upgrade path that does not require third-party agent builders or custom prompt engineering.
Evergreen Goodwill and Project C.U.R.E. show what nonprofit AI deployment actually looks like in 2026 in Geekwire feature
A Geekwire feature published in 2026 profiles two real-world nonprofit AI deployments. Seattle's Evergreen Goodwill is using a Microsoft AI for Good Lab $5M Azure-grant award to classify, sort and price 45 million donated items annually with a human-in-the-loop pricing model. Denver medical-aid charity Project C.U.R.E. is applying AI to predict supply needs and search its operating manuals database for specific fixes. The piece frames sector mood as enthusiasm-mixed-with-anxiety as donors scrutinise AI deployment alongside immediate operational gains.
geekwire.com ↗Why it matters
Both case studies show frontline nonprofit AI is arriving via existing cloud-vendor partnerships rather than fresh tool searches, and any AIH member already running Microsoft 365 or AWS should treat their incumbent vendor's AI-for-Good arm as the first procurement conversation, not a parallel one.
Sentient Futures launches AI×Animals Fellowship as Bay Area animal-welfare movement formally moves to recruit AI
Sentient Futures, the rebranded organisation focused on AI-driven animal-welfare work, announced a spring 2026 AI×Animals Fellowship described as an 8-week course preparing learners for impactful careers at the frontier of AI and animals. The launch followed a February gathering at Mox in San Francisco where animal-welfare advocates and AI researchers met to map applications across wildlife conservation, factory-farming oversight and AI sentience itself. MIT Technology Review framed the moment as "the Bay Area's animal welfare movement wants to recruit AI."
forum.effectivealtruism.org ↗Why it matters
This is the first dedicated talent pipeline for AI-applied animal-welfare work, and any animal-rights or sanctuary nonprofit watching the AI shift now has a concrete fellowship route to seed staff and a peer community to join, rather than building everything from scratch.
Anthropic publishes "AI Fluency for Nonprofits" free training course aimed at nonprofit professionals running Claude in mission delivery
Anthropic launched its AI Fluency for Nonprofits course on its Skilljar platform, designed to help nonprofit professionals develop AI fluency for organisational impact while preserving mission and values. The course covers prompting, workflow design, governance and deployment patterns specific to nonprofit operating models, and is free to access without enterprise procurement. It complements the Anthropic free-tools partnership announced through NTEN's AI Tech Readiness Cohort.
anthropic.skilljar.com ↗Why it matters
This closes the gap between nonprofits with $0 training budget and the production-grade AI deployment patterns that paid Anthropic enterprise customers receive, and any AIH member onboarding staff to Claude now has a vendor-built curriculum to point them at instead of generic ChatGPT tutorials.