Deep Research is now available to all users for free on Gemini
New Model Releases
AI2 Releases OLMo 32B: AI2 has released OLMo 32B, a fully open-source AI model that outperforms GPT 3.5 and GPT 4o mini. The release includes all artifacts, such as training code, data, model weights, and reproduction instructions, allowing for full control and customization.
Sesame Releases CSM 1B Voice Cloning Model: Sesame has released their 1B CSM, a conversational speech model that generates RVQ audio codes from text and audio inputs. The release includes installation steps and a demo on Hugging Face, but parts of the pipeline are missing, leading to community speculation and criticism.
NousResearch Releases DeepHermes Models: NousResearch has released two new AI models: DeepHermes-3-Mistral-24B and DeepHermes-3-Llama-3-3B. These models are available on Hugging Face and feature improved instruction-following capabilities and faster performance compared to their predecessors.
New Services and Features
Gemini Offers Free Deep Research Feature: Gemini has made its Deep Research feature available to all users for free, allowing users to perform a limited number of deep research tasks per month. This feature acts as an AI agent that compiles reports based on live, up-to-date web data, making it useful for tasks that require current information.
Google DeepMind Upgrades Deep Research with Gemini 2.0 Flash Thinking: Google DeepMind has announced an upgrade to Deep Research in the GeminiApp with Gemini 2.0 Flash Thinking Experimental, allowing all users to try this feature at no cost. The new feature is noted for having the lowest hallucination rate among current models on the market.
Claude 3.7 Reasoning Model Available on Android: The Claude 3.7 reasoning model is now available on Android. Users have reported that the update is rolling out in batches, and some features like the model selector are not yet available.
Benchmarking and Performance Updates
Qwq-32b Model Updates and Performance Benchmarks: The Qwq-32b model has been updated on Livebench, showing improved performance and creative capabilities. Users report positive experiences, and the model is being used for various tasks, including search and educational purposes. Benchmark results compare various AI models, with a focus on the Qwq-32b model by Alibaba, which outperforms DeepSeek R1 in several performance metrics.
Major Announcements
Google Solves One-shot Character Consistency: Google has made significant progress in one-shot character consistency, which is crucial for creating consistent characters in AI-generated content. This advancement is notable for its potential applications in gaming and other creative industries, although users have noted that the technology is still not perfect and has some limitations.