Top AI News: Mistral, OpenAI, LG, Gemma, 3D Model and more!

Key Concepts

Gemma 3, Mix Small 3.1, XA1, Hunan (3D generation, T1), Ernie 4.5, X1, Command-R+, Gemini 2.0 Flash, Cosmos Transfer One, Roblox 3D generation, Stability AI multi-view video generation, Open Manners AI, Fish CAD, Anthropic web search, OpenAI speech-to-text & text-to-speech, KBLAM, Edit image in Gro, Deeper Search, Mirage, Mind Maps in NoteLM, Claude 3.5 Sonnet on Cursor, Wave 5 from Vinsurf, Jama package.

Model Releases

• Gemma 3: LG's most capable model for single GPU use.
  • 27 billion parameters.
  • Outperforms DeepSeek B3.
  • Pre-trained on 140 languages.
  • Multimodal capabilities.
  • 128,000 token context length.
• Mix Small 3.1: Multimodal, multilingual, Apache 2.0 licensed, fully open source.
  • 128,000 context length.
  • High performance and faster inference.
  • Better than Gemma 3 (27B parameter model) and GPT4-mini.
• XA1 (LG): Deep reasoning model specialized in math, science, and coding.
  • Three versions: 32B, 7.8B, and 2.4B parameters.
  • Performance on par with DeepSeek R1.
• Hunan 3D Generation Model: Two versions (MV and Mini).
  • Open source.
  • Significant improvement over previous models, even the Mini version.
• Hunan T1: Hybrid Mamba-Transformer architecture.
  • Longer context length, strong logic, concise writing, low hallucination, blazing fast, excellent long text processing.
  • Performance comparable to DeepSeek R1 and 01, and beats GPT 4.5.
• Baidu Ernie 4.5 & X1:
  • Ernie 4.5 outperforms GPT 4.5 at 1% of the price.
  • X1 matches DeepSeek R1 at 50% lower cost.
  • Deep thinking multimodal model.
• Command-R+: Max performance, minimal compute.
  • Optimized for speed, efficiency, and security.
  • Outperforms GPT-4-0 and DeepSeek version 3.
  • 256,000 context length.
  • Advanced RAG, secure and private.
  • 23 supported languages.
• Gemini 2.0 Flash: Native image generation, conversational editing, text in images.
  • Optimized for chat.
  • Example: Adding white and pink flowers to a table in an image.
• Cosmos Transfer One: Conditional world generation with adaptive multimodal control (Hugging Face).
  • Supports RGB camera, depth camera, and HD map data.
  • Creates photorealistic images from various sensor data.
  • Offers different variations like night, foggy night, rainy, snow, and daytime.
• Roblox 3D Generation: Generates 3D objects directly within Roblox by typing a description.
• Stability AI Multi-View Video Generation: Creates 3D videos from 2D images with 3D camera control.

Other Releases and Updates

• Open Manners AI: Automates complex tasks.
• Fish CAD: AI-powered text-to-3D model generation directly in the browser.
  • Example: Creating a 3D model of "woolly" from text.
• Anthropic Web Search: Web search integration for Claude, providing relevant information for questions. Competes with OpenAI and Perplexity.
• OpenAI Speech-to-Text & Text-to-Speech:
  • Speech-to-text: GPT-4-0 transcribe and GPT-4-0 mini transcribe, outperforming previous Whisper versions.
• Microsoft Research KBLAM: Embeds knowledge in a model without retraining.
  • Ingests custom data (e.g., codebase) for information retrieval.
  • Comparison with RAG (external retriever) and in-context learning. KBLAM uses an internal retriever.
• Edit Image in Gro: Image modification by uploading an image and providing instructions. Similar to Gemini image editing.
• Deeper Search (Gro): Extended search and more reasoning capabilities compared to the previous Deep Search.
• Mirage: Creates high-converting ads with virtual spokespersons.
  • Customizable avatars and scripts.
• Mind Maps in NoteLM: Generates mind maps from uploaded documents for better content understanding.
• Claude 3.5 Sonnet on Cursor: Integration with Cursor for extended thinking and increased intelligence.
• Wave 5 from Vinsurf: Advanced version with automatic prediction of typing, error correction, and module import.
• Jama Package: Fine-tuning models with Lora, PFT, multimodality, and tokenization.
  • pip install jama for installation.

Synthesis/Conclusion

This week saw a flurry of AI model releases and updates, spanning various modalities and applications. Key highlights include the release of powerful open-source models like Gemma 3 and Mix Small 3.1, advancements in 3D generation with Hunan and Stability AI, and improvements in reasoning and performance with models like XA1, Ernie 4.5, and Command-R+. Furthermore, tools like Fish CAD, Anthropic Web Search, and Mirage are democratizing AI by making it easier to create 3D models, access information, and generate marketing content. The integration of AI into existing platforms like Roblox, Cursor, and NoteLM demonstrates the increasing pervasiveness of AI across different domains. The development of techniques like KBLAM for knowledge embedding and the Jama package for fine-tuning further empower users to customize and leverage AI models for specific needs.