LLM News

TIIUAE has released Falcon Perception, a 0.6B-parameter early-fusion Transformer that combines image patches and text in a single sequence for open-vocabulary object grounding and segmentation. The model achieves 68.0 Macro-F1 on SA-Co (vs. 62.3 for SAM 3) and introduces PBench, a diagnostic benchmark that isolates performance across five capability levels. TIIUAE also released Falcon OCR, a 0.3B model reaching 80.3 on olmOCR and 88.6 on OmniDocBench.

Google is rolling out AI Inbox in Gmail to AI Ultra subscribers ($249.99/month) following its January announcement. The feature uses Gemini 3 to automatically surface actionable tasks and topics, replacing the need to manually scan message lists.

A source code leak in Anthropic's Claude Code 2.1.88 update exposed more than 512,000 lines of TypeScript, revealing unreleased features including a Tamagotchi-like pet interface and a KAIROS feature for background agent automation. Anthropic confirmed the leak was caused by a packaging error, not a security breach, and has since fixed the issue.

Amazon Bedrock AgentCore Evaluations, a fully managed service for assessing AI agent performance, is now generally available following its public preview debut at AWS re:Invent 2025. The service addresses the core challenge that LLMs are non-deterministic—the same user query can produce different tool selections and outputs across runs—making traditional single-pass testing inadequate for reliable agent deployment.

Google is rolling out a new glow effect to the Gemini overlay on Android, with the animated multicolor border now appearing around the entire screen perimeter instead of just the prompt box. The company has also repositioned Temporary chat to the top-right corner on web, making it more prominent while reducing side panel clutter.

xAI has released Grok 4.20 Multi-Agent, a variant designed for collaborative agent-based workflows with a 2-million-token context window. The model scales from 4 agents at low/medium reasoning effort to 16 agents at high/xhigh effort levels, priced at $2 per million input tokens and $6 per million output tokens.

xAI released Grok 4.20 on March 31, 2026, its flagship model featuring a 2 million token context window, $2 per million input tokens and $6 per million output tokens pricing, and toggleable reasoning capabilities. The model includes web search functionality at $5 per 1,000 queries and claims industry-leading speed with low hallucination rates.

AWS has released QA Studio, a reference solution for QA automation built on Amazon Nova Act that enables teams to define tests in natural language rather than code. The system uses visual understanding to navigate applications like users do, automatically adapting to UI changes and eliminating maintenance overhead from traditional selector-based testing frameworks.

Amazon has added conversational food ordering to Alexa+, its next-generation AI assistant, enabling users to order from Uber Eats and Grubhub through natural language. The feature rolls out today to Alexa+ customers with Echo Show 8 devices and larger, allowing users to browse menus, customize meals, and modify orders mid-conversation.

Google announced Veo 3.1 Lite, a cost-reduced video generation model priced at less than 50% of Veo 3.1 Fast's cost. The model supports text-to-video and image-to-video generation at 720p or 1080p resolution with customizable durations of 4s, 6s, or 8s, rolling out today on the Gemini API and Google AI Studio.

GitHub's Applied Science team deployed coding agents to automate parts of their own development workflow, testing how AI agents can handle increasingly complex programming tasks. The experiment reveals practical insights into agent-driven development patterns and limitations.

IBM announced Granite 4.0 3B Vision, a 3 billion parameter vision-language model designed for enterprise document processing. The model achieves 86.4% on Chart2Summary and 92.1% TEDS score on cropped table extraction, shipped as a LoRA adapter on Granite 4.0 Micro to enable modular text-only fallbacks.

Google has completed a wide rollout of Ask Maps, a Gemini-powered conversational feature in Google Maps, to all users in the US and India as of this week. The feature answers complex, real-world location-based questions that traditional map search cannot handle, such as finding romantic date spots, scenic running routes, and personalized travel recommendations.

Alibaba has released Qwen3.5-Omni, an omnimodal model handling text, images, audio, and video with a 256,000-token context window. The model reportedly outperforms Google's Gemini 3.1 Pro on audio tasks with support for 74 languages in speech recognition, a 6x increase from its predecessor. An unexpected emergent capability: writing working code from spoken instructions and video input, which the team did not explicitly train.

OpenAI released a plugin that embeds its Codex coding assistant directly into Anthropic's Claude Code, the market-dominant code IDE. The plugin offers standard code review, adversarial review, and background task handoff capabilities, requiring only a ChatGPT subscription or OpenAI API key.

Microsoft released the Harrier-OSS embedding model family, comprising three variants with 270M, 600M, and 27B parameters. The largest model achieves 74.3 on the Multilingual MTEB v2 benchmark. All models support 32,768 max tokens and multilingual inputs across 40+ languages.

Google has released Lyria 3 Clip Preview, a music generation model available through the Gemini API as of March 30, 2026. The model generates 30-second audio clips from text prompts or images at $0.04 per clip, with a 1,048,576 token context window.

Google has released Lyria 3 Pro Preview, a music generation model capable of producing full-length songs with verses, choruses, bridges, vocals, and timed lyrics from text prompts or images. The model features a 1,048,576 token context window and charges $0.08 per generated song through the Gemini API.

GitHub has disabled Copilot's ability to insert promotional tips into pull requests following developer backlash. The feature, which injected Raycast ads into over 11,400 pull requests without explicit developer consent, was enabled when Copilot was mentioned in PRs it didn't create. GitHub's product manager acknowledged the decision was a mistake.

Microsoft has released Harrier-OSS-v1, a family of multilingual text embedding models trained with contrastive learning and knowledge distillation. The 0.6B parameter variant achieves a 69.0 score on the Multilingual MTEB v2 benchmark with support for 32,768 token context windows and 45+ languages.