1. Veo 3.1 Lite: Lower-Cost Video Generation via Gemini API
Google introduced Veo 3.1 Lite, a new generative video model tier available through the Gemini API. The model delivers the same generation speed as Veo 3.1 Fast but operates at under half the cost. This pricing structure is designed to support high-volume, production-scale video generation applications where cost per second has previously been a bottleneck.
2. TRL v1.0: Hugging Face Stabilizes Post-Training API
Hugging Face has officially released TRL (Transformer Reinforcement Learning) v1.0, transitioning the library into a stable, production-ready framework. The release provides a unified, standardized API for large language model post-training workflows. It supports over 75 post-training methods, including Supervised Fine-Tuning (SFT), Reward Modeling, Direct Preference Optimization (DPO), and Group Relative Policy Optimization (GRPO).
3. GLM-5V-Turbo: Zhipu AI Releases Multimodal Vision Coding Model
Zhipu AI launched GLM-5V-Turbo, a native multimodal vision-language model optimized for generating code from visual inputs like design mockups and screenshots. The model processes images, video, and text to support agentic engineering workflows. It includes native support for tool calling, task decomposition, GUI interaction, and integration with the OpenClaw framework.
4. Storage Buckets for Spaces: Persistent Volumes on Hugging Face
Hugging Face introduced Storage Buckets for Spaces, enabling users to mount persistent storage volumes directly into their deployed environments. Developers can create or select buckets, configure mount paths, and set access modes within the Space settings. This feature facilitates caching model weights, storing user uploads, and sharing files across multiple Spaces within the same organization.
5. Gemini API Docs MCP and Developer Skills: Tools for Coding Agents
Google introduced the Gemini API Docs Model Context Protocol (MCP) and Gemini API Developer Skills. These tools provide coding agents with direct access to the most up-to-date Gemini API documentation and best practices. By mitigating issues caused by outdated training data, the combined tools enable agents to achieve a 96.3% pass rate on Google's evaluation set.
6. LFM2.5-350M: Liquid AI Releases Compact Edge Model
Liquid AI released LFM2.5-350M, a 350-million parameter model built on the LFM2 architecture. The model was trained on 28 trillion tokens and utilizes large-scale reinforcement learning to improve performance. It is specifically optimized for edge deployment, focusing on tasks such as data extraction and tool use.
7. Semi-Formal Reasoning: Meta Publishes Structured Prompting Technique for Code Review
Researchers at Meta introduced "semi-formal reasoning," a structured prompting technique designed to improve large language model performance on repository-scale code review tasks. The method requires the AI agent to explicitly state premises, trace concrete execution paths, and derive formal conclusions in a logical certificate before answering. This approach bypasses the need for computationally heavy dynamic execution sandboxes while reducing unsupported guesses and hallucinations.