Models & Labs
Google DeepMind Launches Gemma 4 12B Multimodal Model
Google DeepMindhigh confidence
Why it matters
- →Gemma 4 12B's encoder-free architecture reduces latency and memory usage.
- →It enables advanced AI capabilities on consumer-grade hardware.
- →The open-source release fosters innovation and accessibility in AI development.
Google DeepMind has introduced Gemma 4 12B, a new multimodal AI model that integrates vision and audio inputs directly into its language model backbone, eliminating the need for separate encoders. This design reduces latency and memory usage, allowing the model to run on consumer laptops with 16GB of RAM. Released under an Apache 2.0 license, Gemma 4 12B is accessible to developers for building advanced AI applications. The model's performance approaches that of larger models, making it a versatile tool for multimodal and agentic tasks.
Read originalMore from Google DeepMind
© Google DeepMindModels & Labsmodels
Google DeepMind unveils DiffusionGemma for faster text generation
Google DeepMind's DiffusionGemma marks a significant shift in text generation by leveraging diffusion techniques to generate text blocks up to four times faster than traditional models. This 26B Mixture of Experts model, designed for speed-critical applications, moves beyond the sequential token-by-token approach, allowing for parallel generation of 256 tokens. While it offers blazing fast inference on GPUs, it trades off some quality compared to the standard Gemma 4 models. This innovation is particularly beneficial for developers working on real-time interactive AI applications, as it maximizes hardware utilization and reduces latency bottlenecks.
Google DeepMind
© Google DeepMindModels & Labsmodels
Google DeepMind launches Gemini 3.5 Live Translate
Google DeepMind's Gemini 3.5 Live Translate marks a significant leap in real-time speech translation, offering fluid and natural-sounding translations across 70+ languages. Unlike traditional systems, it provides continuous translation, maintaining the speaker's intonation and pacing, and operates just seconds behind the speaker. This model is now available for developers via the Gemini Live API and is being integrated into Google Meet and the Google Translate app. The rollout promises to enhance multilingual communication in various settings, from business meetings to everyday conversations.
Google DeepMind
© Google DeepMindResearchresearch
AI Enhances Learning in Sierra Leone Study
Google DeepMind's recent study in Sierra Leone demonstrates the potential of AI as a powerful educational tool, enhancing rather than replacing traditional teaching methods. The trial showed significant improvements in students' math scores, with AI-driven Guided Learning fostering deeper understanding rather than rote solutions. Teachers reported professional growth, shifting from lecturers to facilitators, as they integrated AI into their lessons. This approach not only increased student engagement but also shifted their focus towards skill-building. The study's success suggests a promising future for AI in education, with plans to expand trials globally.
Google DeepMind
More in Models & Labs
© The Verge AIModels & Labsmodels
Anthropic's Claude Fable 5 Limits Biology Queries
Anthropic's release of Claude Fable 5, touted as their most powerful AI model, comes with significant limitations in answering biology-related questions. This is due to the model's conservative safeguards designed to prevent misuse in bioweapons research. While the model excels in cybersecurity tasks, its biology filters are so stringent that even basic queries like 'what are mitochondria' are blocked. Anthropic aims to balance safety with utility, promising future adjustments to reduce false positives and potentially open up more capabilities for scientific research.
The Verge AI
© GitHub ChangelogModels & Labsmodels
Claude Fable 5 now available in GitHub Copilot
Claude Fable 5, a new model from Anthropic's Mythos class, is now integrated into GitHub Copilot, offering enhanced capabilities for long-horizon coding and knowledge tasks. This model stands out by requiring data retention for safety purposes, a shift from the zero data retention policy of previous models. It promises more efficient coding workflows with fewer tool calls and lower token consumption. Available to select GitHub Copilot users, this rollout marks a significant step in autonomous coding, though it comes with new data handling considerations.
GitHub Changelog
© Hugging Face BlogModels & Labsmodels
Cohere Launches North Mini Code Model for Developers
Cohere has unveiled North Mini Code, a 30B-parameter Mixture-of-Experts model designed for complex software engineering tasks, now available on Hugging Face. This model stands out with its agentic coding capabilities, optimized for terminal-based tasks and high-quality code generation. It outperforms several larger models in coding benchmarks, showcasing its efficiency and robustness. By employing a unique training approach with supervised fine-tuning and reinforcement learning, North Mini Code aims to serve as a reliable foundation for coding agents. This release marks a significant step in making advanced coding models accessible to developers.
Hugging Face Blog