Agent Mode News

uses some non-AdamW optimizer

>= 50% on TerminalBench (https://www.tbench.ai/leaderboard)

Some variation of NSA (Native Sparse Attention)

>= 73% on SWE-Bench Verified (according to epoch.ai)

>= 2 shared experts

>=25T pretraining tokens

>=52B active parameters

>=1.5T parameters

>=16 active experts

1M+ Context

Some image input (multimodality)

>=512 experts

intra-expert communication

DS-MoE with adaptative expert count

>= 60% on BrowseComp (https://www.kaggle.com/benchmarks/openai/browsecomp)

Gemini 2.5 Pro tier or higher on FictionBench (90.6%+ at 192k)

>= 44% on Humanity's Last Exam (text only) at scale.com leaderboard

DeepSeek reports some results with a full-blown deep research agent, and emphasizes that this is the intended use-mode

NO

YES