Hands-on reviews of the AI tools that actually work for work, content, and code. No hype, just what we tested.
Speech recognition is becoming a language problem, not just an audio one. Tencent Hunyuan’s Hy ASR 3.0 preview, built on the Hy3 large language model with a mixture-of-experts architecture, is the company’s version of that shift, recognition that understands context. The numbers on open benchmarks are strong. The model reports a word error rate of … Read more
One API call now covers video, audio, image, and motion prediction. FLUX 3 Video from Black Forest Labs is available on OpenRouter for everyone, according to the platform’s announcement, as a unified multimodal model family trained jointly on a single architecture. That is a different shape from the typical video model release. Most video generators … Read more
Alibaba just put the second-strongest image model, following the Qwen3.8 release, on the Arena leaderboard behind a cloud API. Qwen-Image-3.0-Pro and a Standard variant are now live on Qwen Cloud, with the Pro model ranking first among Chinese models and second overall. The specs are aimed at practical image work. The model handles 4.5k-token prompts, … Read more
Five days after Anthropic doubled its flagship price, OpenAI cut GPT-5.6 Luna by 80 percent. And the cut is permanent, not a temporary promotion, according to Tibo, an AI industry commentator who pushed back on the reading that it was a short-term stunt. The context matters. Luna’s price drop came five days after Anthropic priced … Read more
The compute market has a new kind of deal in it. Anthropic signed a $10 billion agreement with Volta, a cloud startup that is only months old, according to reporting shared by AI industry observers. The deal works out to roughly $1.7 billion per year. What makes this notable is not the headline number. It … Read more
Run an 80-billion-parameter Qwen model on a Mac with about 4.3GB of peak memory. That is the claim behind Swiftlet, a new open-source runtime that targets the Qwen3-Next and Qwen3.5/3.6 MoE hybrid model family and also supports iPhones. The key technical idea is expert streaming. Instead of loading the full model into memory, Swiftlet keeps … Read more
The AI video generation market in 2026 is defined by intense competition. ByteDance’s Seedance 2.5, MiniMax H3, PixVerse, Kling, and others are all pushing strong models. The interesting story is not which is best on benchmarks — it is how the open-source vs proprietary split is shaking out. For users, the practical implication is more … Read more
AirLLM is a project that makes it possible to run 70-billion-parameter language models on consumer hardware — specifically, on a single GPU with as little as 4GB of VRAM. The technique is layer-by-layer inference, which loads and processes one model layer at a time rather than the entire model at once. The practical implication is … Read more
According to recent OpenRouter usage data, Xiaomi’s MiMo model has moved into the global number two position for AI model usage, behind only DeepSeek-V4-Flash. The data, cited by 36Kr based on OpenRouter’s weekly figures, shows the top five models in usage are all Chinese. For the AI model market, this is a significant data point. … Read more
Saner.AI, a Chinese brain-computer interface company, has completed a 3.3 billion yuan angel round, the largest such round in the Chinese BCI space to date. The company is focused on invasive BCI products, with funding going to clinical trials and team expansion. The round is a signal that the BCI space in China is getting … Read more