Summary
AI Search delivers a comprehensive weekly roundup covering an unusually dense wave of model and tool releases. On the foundation model side: DeepSeek ships V4 Pro, a 1.7-trillion-parameter mixture-of-experts model with DS-Park speculative decoding that benchmarks competitively against Kimi K3 and Claude Opus on knowledge and agentic coding tasks while undercutting rivals at just 6 cents per task. xAI releases Grok 4.6, ZhipuAI releases GLM 5.3 (framed as the strongest available open-source model at time of recording), Google releases Gemini 3.7 positioned as its fastest model, and Alibaba releases Qwen 3.8 27B as the leading offline-capable mid-size option.
On the media generation side, two open-source video models ship: LTX Video 2.5 (a 16B multimodal diffusion transformer generating 720p at 30+ fps, Apache 2.0 licensed, compatible with existing LTX 2 LoRAs) and Minimax, with the video comparing quality and speed tradeoffs between them. Tencent releases Scope, a camera-path control framework for video generation built on Wan 2.2. A new open-source music generator small enough to run on low-end GPUs is also highlighted, alongside a new state-of-the-art text-to-speech system.
Additional coverage includes OpenAI previewing an ultra-fast mode for GPT-5.6 Soul at up to 750 output tokens per second (14x standard speed), and Google DeepMind releasing a sign language-to-text model trained on 100,000+ hours across 50+ sign languages, launching initially with American Sign Language on Pixel 11 via Gboard and Live Transcribe.
📺 Source: AI Search · Published August 16, 2026
🏷️ Format: Roundup
