Deepseek OCR 2 cuts visual tokens by 80% and outperforms Gemini 3 Pro on document parsing
Deepseek has unveiled a vision encoder that processes image information based on meaning rather than position. The approach uses far fewer tokens and improves document recognition.
AI coding assistant Cursor now indexes large codebases in 21 seconds instead of over four hours. The trick: instead of building an index from scratch for each new user, Cursor reuses existing indices from team members. According to the company's blog post, copies of the same codebase within a team are 92 percent identical on average, making this approach highly efficient.
Merkle trees compare file hashes between client and repository, only synchronize files that differ and delete missing entries.
A Cursor study found that the semantic search enabled by these indices improves AI response accuracy by 12.5 percent. The technology relies on Merkle trees - a data structure using cryptographic hashes - to ensure users only see code they're authorized to access. For typical projects, wait times for the first search query drop from nearly 8 seconds to just 525 milliseconds. The startup behind Cursor shipped version 2.0 with its own coding model in October 2025 and now generates around $500 million in annual revenue.
Google is working on a feature that lets users clone their own voice in AI Studio. According to TestingCatalog, a hidden option called "Create Your Voice" shows up when selecting the "Flash Native Audio Preview" model, which is currently tied to Gemini 2.5 Flash. Clicking it opens a window for recording and uploading audio, but the feature isn't functional yet. The discovery suggests Google is getting ready to ship native audio capabilities with Gemini 3 Flash. This would let developers create artificial voices based on recorded voice samples. Google released an update for Gemini 2.5 Flash Native Audio back in December 2025 that improved voice quality and made the model follow instructions more precisely.
The hidden "Create your voice" option in Google AI Studio hints at upcoming voice cloning functions.
In addition, a new option has been found that allows entire code collections to be imported via GitHub repositories. The start page is also apparently being revised and will display activities and usage statistics separately in future.
Moonshot AI releases Kimi K2.5, claims most powerful open-weight model with 100-agent coordination
Moonshot AI has released Kimi K2.5, which the company says is the most powerful open-weight model available. The model can independently coordinate up to 100 AI agents working in parallel on complex tasks.
The British government has chosen Anthropic to develop an AI assistant for theGOV.UKwebsite. The Department for Science, Innovation and Technology (DSIT) plans to use the system to help citizens navigate government services and receive personalized guidance. The initial focus will be on jobseekers - helping them with career advice, connecting them to training opportunities, and explaining available programs.
The partnership builds on a declaration of intent signed in February 2025. Anthropic engineers are collaborating directly with UK officials to ensure the government can eventually run the system on its own. Users will keep full control over their data and can opt out at any time.
Anthropic's regional head Pip White said the collaboration demonstrates how AI can be deployed safely for the public good. The company isn't the only US tech firm making moves in the UK - Microsoft, OpenAI, and Nvidia committed over 31 billion pounds to British AI infrastructure last year.
