LLMs

A new framework enables LLM agents to perform reasoning directly in latent space while using text for external actions. This approach decouples internal cognitive processing from the constraints of token-by-token text generation. The method aims to improve efficiency and performance in complex agentic tasks by bypassing verbose intermediate steps.

• Reasoning occurs in latent vectors, not generated text tokens
• External interactions and actions remain standard text-based
• Potentially reduces token costs and latency for complex agent workflows
• Code and models available via Together Computer repositories

Open-weight models now account for 71.5% of token processing on the Vercel AI Gateway. Despite this dominant share of volume, they represent only a quarter of estimated spending. This disparity suggests these models are significantly more cost-efficient than proprietary alternatives.

• Open-weight models dominate Vercel gateway traffic with 71.5% share
• Proprietary models still command 75% of estimated spend despite lower volume
• Open-weight adoption driven by cost efficiency, not just capability
• Practitioners can reduce AI infrastructure costs by shifting to open-weight

Hugging Face has introduced Beam, a new open-weight model with 501 billion parameters. The announcement highlights the model's capacity for ownership, customization, and scaling.

• Beam is a 501B parameter open-weight model
• Designed for local ownership and customization
• Focuses on scalability for large-scale deployments
• Released via Hugging Face Blog

DeepSeek has released a new search and fetch API capability priced at $0.75 for every 1,000 successful requests. This offering provides a cost-effective option for integrating real-time web data retrieval into AI applications. The pricing structure targets high-volume usage scenarios where budget efficiency is critical for scaling.

• New API endpoint enables web search and content fetching within DeepSeek workflows
• Pricing set at $0.75 per 1,000 successful requests, competitive for high-volume apps
• Cost model applies only to successful requests, potentially reducing waste from failures
• Integration allows AI agents to access current web data without separate scraping tools

Google has introduced Gemini 4 Argon, a new model variant designed to tackle difficult computational and reasoning tasks. The announcement emphasizes its capability to handle complex problems, though specific technical details remain limited in the provided excerpt.

• New Gemini 4 Argon model targets hard problem-solving scenarios
• Release focuses on advanced reasoning capabilities
• Limited technical specifics available in current announcement