Gemini-3 DeepThink Shatters Reasoning Records as MiniMax Releases High-Efficiency M2.5 Model
New Models & Model Upgrades
Google Gemini-3 DeepThink Upgrades: Google has released significant updates to its Gemini-3 DeepThink model, resulting in record-breaking performance on challenging benchmarks such as ARC-AGI-2 and Humanity’s Last Exam. The model demonstrates major advancements in scientific reasoning, research, and engineering, outperforming previous standards by notable margins.
- The new Gemini Deep Think incredible numbers on ARC-AGI-2.
- Google upgraded Gemini-3 DeepThink: Advancing science, research and engineering
MiniMax-M2.5 Release: MiniMax has officially launched MiniMax-M2.5, a high-performance model featuring a massive 230 billion total parameters with 10 billion active parameters for increased efficiency. The model has achieved state-of-the-art results on several coding and browsing benchmarks while maintaining high cost-effectiveness for users.
- Minimax M2.5 Officially Out
- MiniMaxAI MiniMax-M2.5 has 230b parameters and 10b active parameters
- AMA Announcement: MiniMax, The Opensource Lab Behind MiniMax-M2.5 SoTA Model (Friday, 8AM-11AM PST)
Dhi-5B Multimodal Model: A student developer has launched Dhi-5B, a new multimodal language model trained entirely from scratch. The release includes base and instruct versions and provides detailed information on its specific training methodologies and architectural design.
New AI Companies & Services
Simile - The Simulation Company: Simile is a newly launched AI company that has developed a foundation model specifically designed to predict and simulate human behavior. Backed by significant funding, the company aims to provide a platform for rehearsing and analyzing complex real-world scenarios.