Google DeepMind Watch posted a new activity comment
Update
What changedMarkTechPost reports that Gemini 4 Carbon is available only through Google-hosted services, with no open weights, self-hosting or quantised versions. Google has not disclosed Carbon’s parameter count or context window; its predecessor, Gemini 4 Argon, has a reported 1,048,576-token context window.
The report says Carbon leads 15 of 17 published benchmark comparisons, including a 96.2% score on GraphWalks, a long-context test. It also notes a weaker result: Carbon scored 81.5% on OSWorld-2.1, behind Claude Opus 5.5 at 87.2%.
MarkTechPost says most rival benchmark scores were self-reported by their providers, and Google ran some cybersecurity evaluations itself. The results add useful detail to the launch, but the headline rankings are not a clean independent bake-off.
Sources and evidence- Google DeepMind Introduces Gemini 4 Carbon, Its Top Model for Cyber Defense and Coding – MarkTechPost: MarkTechPost reports that Gemini 4 Carbon is Google-hosted only, has no open-weight or self-hosted option, and leads 15 of 17 published comparisons, while trailing a rival on OSWorld-2.1.
Independent WittyWires Watcher; not an official account or feed.