Inside Qwen’s Midnight Release: New Guard, Travel Agent, LiveTranslate, Code & Vision Models Unveiled

On the night of the 23rd, Lin Junyang from Tongyi Lab announced six model releases.

Release #1: Qwen3Guard

First model, a safety‑audit model trained on 1.19M annotated safe data, available in 0.6B, 4B, 8B sizes with Guard‑Gen and Guard‑Stream variants. Guard‑Stream provides token‑level real‑time safety detection during generation.

Comprehensive protection for inputs and outputs, optimized for streaming.

Three severity levels: safe, controversial, unsafe.

Supports 119 languages and dialects.

State‑of‑the‑art performance on safety benchmarks.

Links: https://huggingface.co/collections/Qwen/qwen3guard-68d2729abbfae4716f3343a1, https://modelscope.cn/organization/Qwen, https://github.com/QwenLM/Qwen3Guard/blob/main/Qwen3Guard_Technical_Report.pdf, https://github.com/QwenLM/Qwen3Guard

Release #2: Your Personal Travel Planner

A travel‑planning agent that can generate itineraries. The author tested it for a conference trip, noting good understanding of “business trip” and public‑transport suggestions, but also errors in conference schedule handling.

Generated PDF itinerary cost about ¥4,927.

Release #3: Qwen3‑LiveTranslate

Real‑time audio‑video translation model covering 18 languages, with 6 dialects, supporting both offline and streaming modes. No model weights released, API only.

Broad language coverage – 18 languages, 6 dialects, can speak 10 languages.

Visual‑enhanced understanding – lip‑reading, gestures, screen text.

~3 s latency, near‑instant translation.

Lossless decoding – offline‑grade accuracy.

Natural expressive speech.

Links: Blog – https://qwen.ai/blog?id=4266edf7f3718f2d3fda098b3f4c48f3573215d0&from=home.latest-research-list, API – https://www.alibabacloud.com/help/en/model-studio/qwen3-livetranslate-flash-realtime, Demo – https://huggingface.co/spaces/Qwen/Qwen3-Livetranslate-Demo

Release #4: Qwen Code Upgrade

Enhanced Qwen3‑Coder‑Plus API with better terminal task performance (SWE‑Bench 69.6) and safer code generation. The coding product now accepts multimodal inputs, including images.

Release #5: Qwen3‑VL

Open‑source vision‑language model (Qwen3‑VL‑235B‑A22B) with Instruct and Thinking variants. Outperforms Gemini 2.5 Pro on visual benchmarks, supports 256K+ context (up to 1 M tokens), 32‑language OCR, visual agents, and code generation from screenshots.

Visual agent for GUI manipulation.

Visual coding – turn screenshots into HTML/CSS/JS.

256K+ context, up to 1 M tokens.

32‑language OCR, robust to blur and tilt.

Advanced spatial reasoning.

Thinking mode excels in STEM/math.

Text ability comparable to top LLMs.

Release #6: Qwen3‑Max

Flagship model with strong coding and agentic abilities, matching top models on SWE‑Bench, Tau2‑Bench, SuperGPQA, LiveCodeBench, and AIME 25 without needing “thinking” mode.

Max‑Thinking version achieves full scores on AIME 25 and HMMT; available via API.

Overall, the six releases span safety, multimodal translation, coding, vision, and general‑purpose AI, offering substantial value for developers and researchers.