Meta’s MobileLLM-Pro & Google’s Gemma AI Unlock Cancer Therapy & On-Device Power
New AI Models & Research Breakthroughs
Meta Releases MobileLLM-Pro, a High-Performance 1B Model
Meta introduced MobileLLM-Pro, a 1B-parameter foundational language model that outperforms Gemma 3-1B and Llama 3-1B in pre-training and excels in instruction-tuned tasks like API calling, rewriting, and coding. The model is now available on Hugging Face.
Australian Startup’s Kanon 2 Embedder Outperforms OpenAI & Google in Legal Retrieval
An Australian AI startup’s Kanon 2 Embedder model surpassed OpenAI, Google, and Voyage in legal document retrieval tasks, achieving higher accuracy and faster inference times.
Inference.net’s Schematron: Small Models Rivaling GPT-5 in HTML Extraction
Inference.net launched Schematron, a 3B and 8B model family optimized for structured HTML-to-JSON extraction, delivering 40–80x cost savings over GPT-5 while matching its performance. The models are fully open-source.
Google’s Gemma AI Discovers Novel Cancer Therapy Method
Google’s Gemma AI model identified a new approach to improve tumor responsiveness to immunotherapy, validated through experiments. This marks a milestone in AI-driven scientific discovery.
AI Performance & Benchmarking
Study Reveals Actual Context Windows of Top Models Drop to ~64K Under Stress
Despite advertising 1M and 500K context windows, models like Gemini 2.5 Pro and GPT-5 show effective performance closer to 64K tokens in demanding tests.
AI Engineering & Workflows
Best Practices for Prompt Experimentation & Versioning at Scale
A discussion highlights strategies for managing AI prompts in production, including treat prompts as code, A/B testing, multi-agent tracing, and hybrid human-automated evaluations.
AI Community & Comparative Analysis
Sentiment Analysis Dashboard: Claude Code vs. Codex (Reddit-Based)
A dashboard analyzing Reddit comments reveals Codex is preferred over Claude Code in pricing, performance, and code quality, with visualizations of user sentiment trends.