Summary
AI Search delivers a dense weekly roundup covering an unusually active period across multiple AI verticals. The biggest model news: ByteDance releases Seedance 2.5 (positioned as the current best video generation model), Minimax drops H3 as an open-source video model, and Deepseek releases V4 Flash — a model that reportedly matches GLM and Claude Opus on benchmarks while costing roughly 100 times less, enabled by the new DS-Spark architecture that significantly improves efficiency and throughput over the prior V4 Pro version.
Other notable releases include Netflix’s open-source IDV tool for style-transfer editing (change clothing, background, or lighting while preserving character identity and motion); Crisper Whisper 2, a transcription model family ranging from 0.2B to 2B parameters featuring precise word-level timestamps and both verbatim and cleaned output modes; Google’s latest Gemini Robotics model; and Inkling Small from Thinking Machines, the lab founded by former OpenAI CTO Mira Murati. The roundup also covers a new open-source model trained entirely on AMD chips (not Nvidia), interactive real-time video world models, and an AI object-removal tool called Audiogram.
For practitioners tracking the frontier, this episode is notable for the simultaneous advance across video generation, affordable frontier-class language models, and open-source speech transcription — with several releases offering local or API-accessible deployment.
📺 Source: AI Search · Published August 02, 2026
🏷️ Format: Roundup
