Introducing Structured RAG (S-RAG) S-RAG transforms unstructured data into a structured, query-aware representation. It then uses formal queries over structured data at runtime, so AI21 Maestro can retrieve accurate values, ensure completeness, handle inconsistencies, and
GENERATIVE AI
-
Kimi K2 sets benchmarks as first open-source thinking model since GPT-4
By
–
Kimi K2 isn’t chasing benchmarks anymore, It’s setting them. The world’s first open-source thinking model is here.
And it just made the biggest leap since GPT-4. Try it yourself → http://
kimi.com -
K2 architecture with 1T parameters and 32B active MoE for reasoning
By
–
The K2 architecture? 1T total parameters.
32B active MoE.
Optimized for reasoning, not just response speed. This is what “thinking AI” looks like. -
All open-source, all public: best large model, November 2025
By
–
All open-source.
All public. This is the best open-source and open-access large model you can get as of November 2025. -
A working Obsidian clone built in one prompt, zero errors
By
–
Here are some demos.
— God of Prompt (@godofprompt) 12 novembre 2025
Built a working Obsidian clone in one prompt.
2,000 lines of code.
Zero errors.
Even entry-level engineers can’t do that. pic.twitter.com/1hiw6dD06HHere are some demos. Built a working Obsidian clone in one prompt.
2,000 lines of code.
Zero errors. Even entry-level engineers can’t do that. -
Kimi aced Humanity’s Last Exam, now #1 thinking model
By
–
HLE (“Humanity’s Last Exam”) is brutal thousands of expert-level questions across 100+ subjects. PhD students fail.
GPT-5 struggles.
Kimi aced it. It’s now officially the #1 thinking model in the world. -

Kimi beats Grok-4 on Humanity’s Last Exam with limited budget
By
–
Most AI companies brag about “catching up” to GPT-5. Kimi just skipped ahead. It beat Grok-4 the $5B Elon-backed model on Humanity’s Last Exam, the hardest benchmark ever built for AI. How?
They pulled all that off with barely any budget and super limited resources. -

xAI drops Grok-4-Fast: 2M-token reasoning model, faster, cheaper
By
–
xAI just dropped a monster: Grok-4-Fast Its a 2M-token reasoning model that’s faster than anything in its class and cheaper than lunch. $0.20 in / $0.50 out per million tokens. That’s 6× cheaper than Grok-4-0709 with the same brain. The wild part? It’s not just speed. It’s
-
ElevenLabs Releases Scribe V2 Realtime Transcription Model
By
–
ElevenLabs released Scribe V2 Realtime, a transcription low-latency model with support for 90+ languages. https://t.co/ksVHzto2uI pic.twitter.com/r6QJ1cWTTC
— 🚨 AI News | TestingCatalog (@testingcatalog) 11 novembre 2025ElevenLabs released Scribe V2 Realtime, a transcription low-latency model with support for 90+ languages.
