Model Watch

On June 23, users discovered that ChatGPT quietly launched Bidi 1, a new two-way voice model. Available in the model selector alongside standard and advanced voice modes, its breakthrough feature is real-time 'listen while speaking' — users can interrupt mid-conversation with new instructions. This marks a shift from turn-taking to natural simultaneous conversation. No official announcement yet; wider testing expected this week. A strong response to Google Gemini Live and Anthropic Claude Voice.

Alibaba's Qwen team released Qwen-AgentWorld, a native language world model covering MCP, Search, Terminal, SWE, Web, OS, and Android. The core innovation is 'predict-then-act' — agents simulate actions before executing, reducing trial-and-error costs. Trained on 10M+ real interaction traces through CPT→SFT→RL pipeline. Achieved 58.71 on AgentWorldBench, surpassing GPT-5.4 (58.25) and Claude Opus 4.8. Fully open-source.

Sky Computing Lab released FastWan-QAD, a high-speed video generation series using Quantization-Aware Distillation (QAD) trained on FastVideo. The key selling point: extreme speed — generating a 5-second 480P video in just 1.8 seconds on a single RTX 5090, tens of times faster than traditional methods. QAD lets large models 'teach' smaller models to be leaner while preserving quality. Weights, code, and blog posts are open-sourced for free use.