1. Google Releases Gemma 4 Under Apache 2.0 License
Google has launched the Gemma 4 family of multimodal open-weight models, including 2B, 4B, 31B, and a 26B MoE variant. The models support text, image, audio, and video inputs and feature context windows up to 256k tokens. Notably for enterprise developers, Google has dropped its custom usage restrictions and released the entire family under the standard Apache 2.0 license. The 31B and 26B models can run on a single H100 and show highly competitive reasoning performance against peers like Qwen3.5.
2. Cursor 3 Introduces Multi-Agent Workspace and Cloud-to-Local Handoff
Cursor has released version 3, completely rebuilding its interface to focus on multi-agent software development. The update introduces a unified workspace that allows developers to run multiple local and cloud agents in parallel across different repositories. A new handoff feature enables seamless transitions of agent sessions between cloud environments and local desktops for testing and iteration. The release also adds an integrated browser for prompting against local websites and a marketplace for MCP plugins.
3. Arcee AI Releases Trinity Large Thinking Reasoning Model
Arcee AI has released Trinity Large Thinking, an open-weight reasoning model designed for complex, long-horizon agents and multi-turn tool calling. The model focuses on maintaining coherence across turns and adhering to strict constraints during tool use without degrading output quality. It is available via Arcee's API and as downloadable weights on Hugging Face under the Apache 2.0 license.
4. Research Highlights Prompt Injection Vulnerabilities in LLM-as-a-Judge Workflows
A new report demonstrates that applicants can successfully use prompt injections in documents like CVs and papers to manipulate LLMs acting as automated judges. The testing revealed that while older and smaller models are highly susceptible to these attacks, most current frontier models successfully resist them. However, Gemini 3 was identified as the only frontier-class model vulnerable to this specific type of injection. Developers relying on LLMs for automated evaluation should verify their model's resilience to embedded prompt attacks.
5. IBM Launches Granite 4.0 3B Vision for Document Extraction
IBM has released Granite 4.0 3B Vision, a vision-language model specifically engineered for enterprise document data extraction. Rather than using a monolithic multimodal architecture, the model functions as a specialized adapter attached to the Granite 4.0 Micro language backbone. This design targets high-fidelity visual reasoning for structured data extraction tasks.
6. Fujitsu Open-Sources One Compression Library for LLM Quantization
Fujitsu has released One Compression (OneComp), an open-source Python library for the post-training quantization of large language models. The library implements current quantization algorithms, including GPTQ and DBF. It has been officially verified to work with models like TinyLlama, Llama-2, Llama-3, and the Qwen3 family up to 32B parameters.