15 articles available
The downloadable models return typed decisions without generating tokens, support multimodal inputs, and run on edge hardware—but use a custom license with a commercial revenue threshold.
Read the storyThe MIT-licensed 0.6B and 9B retrievers embed text and page images into a shared multi-vector space, enabling a small query encoder to search a higher-quality 9B-built index.
Read the storyAnthropic’s new small model adds adjustable effort and sharply lower pricing, with broad cloud availability and strong vendor-reported gains over Haiku 4.5.
Read the storyBeat connects to Jira, GitHub, Slack and Granola to prepare reviews, tickets and messages, but requires approval before acting. Limited beta access starts November 1, with paid plans planned for January.
Read the storyMistral’s trillion-parameter multimodal model enters API preview with weights promised later in October. Independent Vals results show strong finance and legal-agent scores but a much more mixed overall profile.
Read the storyAn October 7 blended leaderboard gives Claude Fable 5 an 85–84 edge over GPT-5.6 Sol, but its component boards use different methods and update dates.
Read the story
The Apache-2.0 corpus includes papers, reasoning summaries and partial Lean formalizations, but the underlying model remains private and many results still need independent review.
Read the storyThe new endpoint returns predicates, choices and rubric scores from text or images. OpenAI claims roughly 10× lower latency than the Responses API, but has not published the benchmark methodology.
Read the storyGoogle’s new hosted image model adds stronger text rendering, multi-image consistency and search grounding, while its published context limits conflict across official documentation.
Read the storyThe Apache-2.0 TypeScript toolkit now cancels pending streams and rejects waiting promises on abort, while standardizing provider decision refusals as refusal answers.
Read the story