Google unveils Gemini 4 Carbon, with a benchmark lead and restricted rollout
Carbon tops Artificial Analysis’s Intelligence Index v5. Initial access starts with selected cyber defenders, not a general release.
AI news without the empty calories.
The systems moving AI forward. New models, capabilities, and what the demonstrations really tell us.
Carbon tops Artificial Analysis’s Intelligence Index v5. Initial access starts with selected cyber defenders, not a general release.
Google says students in the United States and India can add a catalog of textbooks to notebooks, with answers grounded in the selected books.
Mistral says its 1-trillion-parameter multimodal model is available to self-host, but businesses serving it as a model service need a separate commercial license.
The enterprise event is confirmed, but the model identities, alleged routing and release dates discussed in an October 29 video remain unverified.
Google expects more models, including efficient versions, but its CEO describes current improvement loops as working with researchers.
A competing model vendor says a sandbox blocked a test-string transfer; intent and root cause remain unproven.
The official model is BF16; a community quantizer reports a heavily compressed configuration for a 12GB GPU.
A dense architecture, quantization work and faster decoding appear in community reports. Alibaba confirms an October 29 conference—not a Qwen 4 release date.
The new model starts at $1 per million input tokens and $5 per million output tokens. OpenAI is pitching faster everyday automation, with stronger coding and visual reasoning.
The redesign moves search, recent chats and account controls into a more compact navigation model, rolling out through a server-side update.
Google is bringing notebook creation, source management and notebook questions into Gemini conversations, with availability and account-region limits still in view.
With permission, Gemini can combine Google Maps searches and saved places with Calendar context, turning personalization into a location-aware recommendation layer.
The research preview puts conversational timing and vocal expression at the center of its design.