Qwen3.8-LiveTranslate interprets 60 languages at a 2.3-second lag
Alibaba's realtime WebSocket model clones each speaker's voice as it interprets — plus GPT-5.5's Oct 14 retirement and an open 149M sparse retriever.
Copy markdown
Simultaneous interpretation is now one WebSocket call
Qwen3.8-LiveTranslate streams speech-to-speech interpretation over a realtime WebSocket (model id qwen3.8-livetranslate-flash-realtime), trailing the speaker by ~2.3 seconds on average — down 18% from 2.8s thanks to a new Interleave architecture. You can stand up a live meeting interpreter or conference-caption tool without duct-taping separate STT, translation, and TTS services together.
It keeps each speaker's own voice, in 29 languages
The model understands 60 languages, speaks 29 of them (English, Chinese, Arabic, Japanese, Hindi and more) and can emit text in 31 others, with speaker diarization and stable voice cloning so a dubbed call still sounds like the person on it. A synchronized bilingual display shows source and target side by side, and it leans on conversation history to keep terminology consistent across a long session.
The catch: closed weights, billed per audio token
There's nothing to download — it's live only on Alibaba Cloud Model Studio and QwenCloud at $7.50 per million audio-input tokens and $30 per million audio-output, roughly $1.54 for an hour of two-way speech. Context caps near 53K tokens with a 10-request-per-minute limit, so plan it into a product rather than a weekend local hack.
GPT-5.5 exits ChatGPT and Codex on Oct 14
OpenAI is retiring GPT-5.5 from ChatGPT, ChatGPT Work and Codex on October 14, routing users to GPT-5.6 Sol or GPT-6 Astra. The raw API is untouched, but Codex users should update workspace defaults, saved model settings, custom agents, scheduled tasks and any script that hardcodes gpt-5.5 before the cutoff.
SPARSEUP puts open sparse retrieval in 149M params
Linkup Research open-sourced SPARSEUP, a 149M-parameter learned sparse embedding model built on ModernBERT under Apache 2.0, with weights already on Hugging Face. It's a fully self-hostable retriever — a trainable cousin of BM25 — you can drop into a RAG stack this weekend with no API bill and no rate limit.