350m5xaccuracyacrossadditionaddsagentagenticagentsanthropicattentionbuildclaudecodecodingcompletecudadeepdriveneveryfoundationframeworkgeminigooglegptgpuharnessinferenceinteractionintroducesk2kernelskimilabslanguagelauncheslearninglfm2llmlocalmeetmemorymetamicrosoftmodelmodelsmoemultimodalnativenvidiaopenopenaiparameterpipelineprimeprotocolpythonrealtimereasoningreleasesreportingresearchrlscaleselfsourcesourcessparsespecsystemthreetooltrainingtutorialusingvisionwebwithout