AI Product Updates: New Features and Releases

Alibaba's real-time speech translation model uses a Hybrid MoE that selects expert modules per task, with audio and translations interleaved in a single processing stream.

The model listens, translates and speaks at once, separating speakers while preserving their voices, returning source and translated text together, supporting 60 audio input and 29 speech output languages, and cutting latency from 2.8 to 2.3 seconds.

In multi-speaker turn-taking conversations, the model labels each utterance with spk1, spk2 and similar tags and keeps each speaker's voice in the translated speech.

More than ten AI projects launched on Base this week, spanning private AI, agent tooling, onchain payments and paid model access.

Notion now offers two GLM open-weight models, GLM 5.3 and GLM 5.3 Flash, in Personal and Custom Agents.

The recipe adds 262K-context throughput numbers on one DGX Spark, NVIDIA's official NVFP4 checkpoint support, 24/7 supervised restarts, cached-token reporting and safer key handling.

CC, an experimental Google Labs assistant, connects to email, calendar, chat and tasks shared by family members and sends a morning briefing of to-dos and schedules.

The new feeders use a built-in scale and, on pricier models, an AI camera to record how much and when a cat eats, with the top health-monitoring features behind an extra subscription.

Alibaba's Qwen3.8-Omni-Flash takes text, audio and video input with a 1M context, plans and calls tools for tasks like video editing and short-drama translation, and is API-only, not open source.

Tencent's WeChat team open-sourced WeVisDoc in 2B and 4B versions, which turn a full page image into Markdown in one pass, with formulas as LaTeX and tables as HTML, without separate OCR or layout detection.

Movit's beta app lets users find travelers already heading to their destination as delivery agents, browse routes and pricing, and switch between sender and delivery agent roles.

Runable now offers Muse Spark 1.3, with roughly 8x cheaper input, about 25% fewer tokens than Spark 1.2, a 1M-token context, and a focus on coding and long-horizon agents.