AI news5
Kimi K3.1 Pro arrives with a one-million-token context
Moonshot has released Kimi K3.1 Pro across its apps, coding tools and API, alongside a Highspeed variant. The model has a one-million-token context window; Moonshot says full weights will follow by 4 November. It describes the model as a post-trained successor to Kimi K3, with improvements in coding, 3D creation and financial research.
kimi.com ↗Why it matters
This new post-training release advances the older K3 model already covered, while its promised weight release remains a future milestone.
Tencent releases WeVisOne-VL for boxes, text and pixel masks
Tencent has introduced WeVisOne-VL in 2B and 4B versions, with a paper submitted on 22 October. It combines object detection, counting, OCR, segmentation and matting in one vision-language model. The project also supports an explicit absent-object response instead of always returning a location.
tencent.github.io ↗Why it matters
This new perception model is distinct from earlier Tencent infrastructure reporting and could simplify tools that need both document text and precise image regions.
Codex CLI 0.164 adds ten reusable prompt slots
OpenAI released stable Codex CLI 0.164.0 on 23 October. Its new /presets command stores prompts and slash commands in ten numbered slots, while inline commands select models, reasoning effort and permissions. The release also fixes failed-compaction history and restores saved settings when resuming sessions.
github.com ↗Why it matters
This specific CLI release improves repeatable agent work and recovery, a separate development from the earlier Codex cloud-distribution story.
Anthropic brings full Claude Platform access to Google Cloud
Anthropic made its managed Claude offering on Google Cloud generally available on 22 October. Customers can use Google Cloud IAM and billing, retire existing cloud commitments, and receive Claude Platform features and betas as they launch. US and Global endpoints are available, with Anthropic operating the service and acting as data processor.
claude.com ↗Why it matters
The new Anthropic-managed route adds day-one platform parity, distinct from earlier Claude features and Google Cloud’s separate OpenAI preview.
Cerebras explains a 59-wafer design for a three-trillion-parameter model
Cerebras published a serving-architecture explanation on 22 October using Kimi K3 as an illustrative example. With 4-bit weights and an assumed 256K average request, it maps the model across 59 WSE-3T wafers, each with 44GB of on-chip memory. Weights stay on the wafers while intermediate activations pass between them.
cerebras.ai ↗Why it matters
The new architecture disclosure explains memory and concurrency constraints behind earlier speed announcements, without establishing a new production benchmark.
AI in the nonprofit sector3
Mozilla opens community AI funding of up to US$500,000 per project
Mozilla Foundation opened its Human-Scale AI Studio call on 22 October for technologies already used or piloted with communities. Selected projects receive US$300,000 and a year of support from April 2027, with a further US$200,000 and two years contingent on progress. Seven projects will be supported in total, including three already selected; applications close 21 November at 6:59pm AEDT.
mozillafoundation.org ↗Why it matters
This new studio cohort differs from Mozilla’s earlier democracy and infrastructure programmes by backing specific community-built products through sustained development.
UNDP and Tencent open a five-country youth AI challenge
UNDP and Tencent Charity Foundation launched the Youth AI4SDGs Impact Challenge on 23 October, focusing on Indonesia, Malaysia, the Philippines, Singapore and Thailand. Participants are offered training, mentorship, innovation grants and access to tools including WorkBuddy and Miora. Applications are open, but the announcement does not specify grant amounts or a closing date.
undp.org ↗Why it matters
This new challenge creates a route from youth prototypes to community pilots, distinct from earlier UNDP programmes and Tencent model or infrastructure news.
Tarbell’s US$10 million AI journalism fund remains open until 11 December
Tarbell Center for AI Journalism is accepting institutional proposals under its US$10 million programme announced on 5 October. Grants range from US$50,000 to US$500,000 or more for newsrooms, teams and media ventures covering AI. The round closes on 11 December; Tarbell says it will formally evaluate only proposals it considers very likely to fund.
tarbellcenter.org ↗Why it matters
The still-open call offers nonprofit newsrooms sustained reporting capacity rather than individual story funding, with editorial decisions retained by recipients.