AI news6
DeepSeek opens its search API and updates V4.1 Flash
DeepSeek's 17 October changelog opens its in-house Search and Fetch API and upgrades its model to V4.1-Flash-1017. The company says its search engine handles hundreds of millions of daily requests. Existing deepseek-flash calls reach the updated model without a change to API pricing.
api-docs.deepseek.com ↗Why it matters
The new retrieval service and dated model update add tools for agent builders beyond DeepSeek's earlier preview and fundraising announcements.
Reflection releases Beam's 501-billion-parameter weights
Reflection released Beam's weights under Apache 2.0 on 16 October, following its earlier preview. The model has 501 billion total parameters and 23 billion active. Weights come in BF16, FP8 and NVFP4 formats, alongside a model card, safety report and integrations with more than 40 partners.
reflection.ai ↗Why it matters
This is the downloadable-weights milestone missing from the preview covered earlier, allowing organisations to evaluate and customise Beam on their own infrastructure.
Jais 3 releases six Arabic-focused models, from 0.9B to 375B
The Institute of Foundation Models and Inception42 introduced Jais 3 on 15 October, with six model sizes and Apache 2.0 weights. The family supports Arabic writing and tool-based agent tasks. Context windows reach 512,000 tokens, with a 128,000-token limit for the smallest 0.9B model.
ifm.ai ↗Why it matters
Arabic-first, locally deployable models give teams a way to test language and dialect needs without relying entirely on a hosted general-purpose assistant.
US agencies publish 10 practices for frontier-agent incidents
NIST, NSA and CISA published voluntary incident-management guidance on 16 October. Its 10 practices span preparation, detection, response, recovery and reporting. The focus is significant uninstructed agent actions during training and evaluation that undermine controls or human oversight.
nist.gov ↗Why it matters
This moves beyond earlier model-specific evaluations to an incident playbook for agent developers, while remaining voluntary guidance.
RapidCanvas raises US$29 million for managed enterprise AI
RapidCanvas announced a US$29 million Series B on 16 October, led by Vertex Ventures SEA & India, bringing total funding to US$52.5 million. Its service pairs an AI platform with field engineers who build and maintain applications. The company says more than 75 enterprises made 1.4 million AI-led decisions with its applications so far in 2026.
rapidcanvas.ai ↗Why it matters
The round backs ongoing implementation and operations, highlighting demand for help moving custom AI into production rather than simply providing model access.
Intuit signs agreement to acquire AI billing platform Sequence
Intuit announced an agreement to acquire Sequence on 15 October, with financial terms undisclosed. Sequence automates pricing, billing, collections and revenue recognition for usage-based and other complex contracts. Intuit plans an integrated offering for its mid-market products in the coming months after the acquisition closes.
investors.intuit.com ↗Why it matters
The planned acquisition connects AI finance workflows with mainstream accounting systems, but an integrated product is still a future release.
AI in the nonprofit sector2
RICH Hub opens rolling climate AI intake with potential US$200,000 grants
LDRI's RICH Hub published an expression-of-interest call on 13 October for African-led or Africa-based climate AI innovations at pilot stage or beyond. Selected projects may receive technical support, partnerships and scaling grants up to US$200,000. Applications are rolling, and submitting an expression of interest does not guarantee support or funding.
climatehub.developlocal.org ↗Why it matters
This new climate-specific pipeline gives African teams a route to scaling support, with eligibility and funding still subject to assessment.
Charity Digital: 62% of survey respondents use AI experimentally
Charity Digital's 15 October early survey findings report 84% using AI, 62% using it experimentally and 23% strategically. Respondents rated their own AI skills and leaders' understanding at five out of 10. The article does not disclose a sample size, and the full report is due at the end of October.
charitydigital.org.uk ↗Why it matters
Unlike earlier risk guides and the sector-wide skills report, these new survey results separate experimentation from strategic adoption, pointing to a training and governance gap.