AI That Thinks on a Budget? Google’s New Gemini Just Did That

⚡ Google Releases Gemini 2.5 Flash – AI That Thinks on a Budget

Google just dropped Gemini 2.5 Flash, a hybrid reasoning model that introduces a game changing feature:

🧠 "Thinking Budget" lets users control how much cognitive effort (and compute cost) the model uses.

🔍 Key Features:

• Matches OpenAI’s o4 mini on performance
  
• Strong in reasoning, STEM, and visual AI tasks
  
• Available via Google AI Studio, Vertex AI, and the Gemini App

📌 Ideal for startups building chatbots, extraction tools, or budget sensitive agents.

💻 Gemma 3 Now Runs on Consumer GPUs with QAT

Google’s Gemma 3, a 27B parameter powerhouse, can now run on a desktop GPU like the RTX 3090.

How?

• ✅ New Quantization Aware Training (QAT) cuts memory usage
  
• 🚀 Maintains model quality with lower hardware demand
  
• 🖥️ Democratizes access for indie developers, researchers, and educators

📌 Previously exclusive to H100 class GPUs now accessible to anyone with a high end gaming PC.

🧪 OpenAI’s o3 Caught Hallucinating 2x More Than o1

Third-party testing revealed that OpenAI’s new o3 model:

• ❌ Fabricates events and actions
  
• 📉 Hallucinates at over 2x the rate of o1
  
• 🧾 Even OpenAI’s system card confirms it

📌 While powerful, o3’s hallucination risk could impact mission critical use cases.

🎬 Alibaba’s Wan 2.1 Video AI Allows Frame Controlled Generation

Alibaba dropped Wan 2.1 FLF2V-14B, a 14B parameter open source video model that:

• 🎥 Generates cinematic 720p video clips
  
• 🎬 Uses first and last frame as generation anchors
  
• 🌀 Produces smooth transitions with controllable action flow

📌 A new competitor in open source video gen alongside Runway, Pika, and Kling.

🧬 Profluent's ProGen3 AI Designs Proteins for Antibodies & Gene Editing

Biotech company Profluent released its new ProGen3 models, led by:

• 🧠 A 46B LLM trained on 3.4B+ protein sequences
  
• 💉 Designs antibodies similar to FDA approved therapeutics
  
• 🧬 Created gene editing proteins 50% smaller than CRISPR Cas9

📌 A leap in AI x Biotech smaller, more versatile, and potentially faster therapies.

🔁 DeepMind Proposes “Experiential Learning” AI Like a Human

DeepMind researchers published a groundbreaking proposal:

• 🔄 AI that learns from real world feedback and experiences not just from human created data.

• 🎯 Implication: AI that improves over time by interacting with the environment and getting feedback similar to how humans and animals learn.

📌 This could lay the foundation for truly adaptive and evolving AI agents.

🧠 Final Takeaways

• Google’s Gemini 2.5 Flash makes AI cost aware and controllable
  
• Gemma 3 now runs on desktop GPUs huge for accessibility
  
• OpenAI’s o3 might need refinement after hallucination concerns
  
• Alibaba and Profluent are pushing AI boundaries in video and biotech
  
• DeepMind is redefining how AI learns from data to experience

📌 Don’t Miss a Day of AI Momentum

Bookmark 👉 SoftlabsGroup.com Your go to for daily AI breakthroughs, tech trends, and strategic updates made for professionals.