September 3, 2026
Quick hits from across the AI world in the last 24 hours.
- ⚡ Quick hits
- Open Source Unsloth enabled one-click local execution of its GGUF models through Hermes, including Qwen3.8-27B, Qwen3.8-Flash, and DeepSeek-V4-Flash. X (@UnslothAI) ↗
- AIGC Recraft made V4 Styles available in ComfyUI, turning reusable visual styles into workflow inputs instead of repeatedly written prompts. X (@recraftai) ↗
- Agents AWS demonstrated a Strands agent creating and validating a missing similarity tool before continuing its assigned article-comparison task. X (@awsdevelopers) ↗
- Agents LlamaIndex added Turbo mode to Extract, claiming roughly fourfold faster structured-data extraction from documents. X (@llama_index) ↗
- Agents Bolt launched Visual Edits, allowing users to select and directly modify text, navigation bars, and buttons on the preview canvas. X (@boltdotnew) ↗
- Agents Perplexity released Portable Computer for Linux systems equipped with Nvidia RTX GPUs carrying at least 24GB of VRAM. X (@perplexity_ai) ↗
- Industry ElevenLabs partnered with Genesys to integrate ElevenAgents into enterprise customer journeys alongside Genesys Cloud Agent Copilot. X (@ElevenLabs) ↗
- Agents Krea opened the beta of Krea Agents, a new platform using specialized interfaces and AI agents for professional creative production. X (@krea_ai) ↗
- AIGC Tripo added multi-view input to its P2.0 Preview, generating 3D geometry from front, left, right, and back reference images. X (@tripoai) ↗
- AIGC Reka and Nvidia demonstrated a steerable 30-billion-parameter video model generating continuous 720p output at 24 frames per second. X (@RekaAILabs) ↗
- Models Microsoft released MAI-Transcribe-2 on Foundry, claiming tenfold GPT-Transcribe speed alongside lower cost and higher transcription quality. X (@MicrosoftAI) ↗
- Agents MongoDB introduced a backend integration providing the virtual file system used by LangChain Deep Agents. X (@MongoDB) ↗
- Open Source Weaviate 1.38 introduced a Boost API for query-time rescoring when hard filters would otherwise exclude potentially useful search results. X (@weaviate_io) ↗
- Models MiniMax said MiniMax-M3 powers HUMAIN-M3, which was additionally trained on more than one trillion Arabic tokens. X (@MiniMax_AI) ↗