DEV Community

Cover image for Dec 12, 2025 | The Tongyi Weekly: Your weekly dose of cutting-edge AI from Tongyi Lab
Tongyi Lab
Tongyi Lab

Posted on

Dec 12, 2025 | The Tongyi Weekly: Your weekly dose of cutting-edge AI from Tongyi Lab

#ai #llm #opensource #coding

Hello, builders and researchers,
This week was nothing short of extraordinary for Qwen — a true harvest of research milestones, product breakthroughs, and community-powered innovation.
From multilingual TTS that sounds human to RL methods that train smarter, we’re witnessing the full arc of what open, thoughtful AI can become.

Let’s dive in.

👉 Subscribe to The Tongyi Weekly and never miss a release:
Subscribe Now：https://www.linkedin.com/build-relation/newsletter-follow?entityUrn=7392460924453945345

📣 Model Release & Updates

Qwen3-Omni-Flash (2025-12-01): Smarter, More Human
What's improved:

  • Enhanced multi-turn video/audio understanding - conversations flow naturally
  • Customize your AI's personality through system prompts (think roleplay scenarios!)
  • Smarter language handling + rock-solid support: 119 text languages | 19 speech
  • Voices indistinguishable from humans

🔗 Try it now:

Qwen3-TTS (version 2025-11-27): Voices That Feel Real
We've leveled up on what matters most:

  • More Personalities: Over 49 high-quality voices, from cute and playful to wise and stern. Find your perfect match!
  • Global Reach: Now speaks 10 languages (zh, en, de, it, pt, es, ja, ko, fr, ru) & authentic dialects (Minnan, Wu, Cantonese, Sichuan, Beijing, Nanjing, Tianjin, Shaanxi)
  • Insanely Natural: The rhythm and speed adapt just like a real person. It's uncanny.

🔗 Try it now:

Qwen Code v0.2.2 → v0.3.0: Stream JSON + Global Ready
Two breakthrough features:

  • Stream JSON Support

• --output-format stream-json for streaming output

• --input-format stream-json for structured input

• 3-tier adapter architecture + complete session management

• Endless possibilities for SDK integration, automation tools, CI/CD pipelines!

  • Full Internationalization

• Built-in EN/CN interface + custom language pack extensions

• /language ui zh-CN - One-click UI switching

• /language output Chinese - Set AI output language

• Global developers welcome to contribute your local language packs!

  • Security & Stability Leap Forward

🔗 GitHub

Qwen Learn Mode — Your Personal AI Learning Tutor
In Qwen Learn Mode, Qwen Chat turns information into understanding that actually sticks.Powered by our Qwen3-Max model and grounded in cognitive psychology, it designs a learning path tailored to the way you think.

  • Guides you through Socratic-style dialogue, instead of just giving you answers

  • Adapts to your current level, like a tutor who always works in your optimal learning zone

  • Builds mental scaffolds so you can handle complex logic without feeling overwhelmed

✨ Try Learn Mode

🧠 Research Breakthroughs

Introducing SAPO: A Smoother Path to RL Training
We introduce Soft Adaptive Policy Optimization (SAPO) — a smooth, stable, and highly effective RL method for training large language models.
SAPO replaces hard boundaries with a continuous, temperature‑controlled gate that:

  • Smooth trust‑region behavior → no abrupt gradient drop
  • Sequence-level coherence → align sequence‑level behavior
  • Token-level adaptivity → preserves useful gradients & boosts sample efficiency
  • Asymmetric temperatures → significantly improved stability, esp. in MoE models What does this mean in practice?
  • Longer stable RL runs
  • Higher Pass@1
  • Stronger performance on Qwen3‑VL across math, coding & multimodal tasks

📄 Paper on arXiv

📚 Technical Blog

🧩 Ecosystem Highlights

Model Milestone: Z-Image-Turbo Ranks #1
According to ArtificialAnlys, Z-Image-Turbo now ranks:

  • #1 Open Source Model
  • Top 10 Overall — the only open model on the list With high-fidelity outputs, $5/1k pricing, and full open source, this is generative AI that’s accessible, affordable, and community-driven.

Qwen3-4B: The #1 Base Model for Fine-Tuning
A rigorous benchmark on small language models by distil labs shows:
Qwen3-4B emerges as the #1 base model for fine-tuning, matching or exceeding a 120B teacher model on 7 out of 8 tasks.
If you need maximum accuracy with efficient compute, Qwen3-4B is your starting point.
If you are looking for maximum accuracy with efficient compute, Qwen3-4B is your top choice.

📄 Read the full report

XiYan-SQL: #1 on All Open BIRD-CRITIC Leaderboards
XiYan-SQL is an innovative natural language–to–SQL conversion framework designed to address the performance challenges large language models face in SQL generation tasks.
XiYan-SQL just hit #1 across all open BIRD-CRITIC (SWE-SQL) leaderboards, the real-world SQL diagnostic benchmark from academia + Google Cloud, built from actual database errors and tricky queries.
Why XiYan-SQL matters

  • Not just text → SQL: it diagnoses and fixes failing queries.
  • Handles complex ops (INSERT / UPDATE / DELETE) across messy, multi-dialect DBs.
  • Remains robust even on unseen, out-of-distribution databases. What this means in practice:
  • More reliable SQL debugging in real, production-style environments
  • Stronger robustness for messy and evolving data stacks

✨ Community Spotlights

Community Celebration: WanMuse+ “Heartbeat” Winners Announced
The winners of WanMuse+ Season 2: “Heartbeat” have been revealed.To every creator who showed AI what it means to feel a heartbeat — we see you, we honor you, and we’re inspired by you.

🎉 Congratulations to all finalists and winners!

🔗 Learn More

Light Migration LoRA: Qwen-Edit-2509-Light-Migration from dx8152
Say goodbye to unnatural lighting artifacts.This Light Migration LoRA from dx8152 for Qwen-Image-Edit-2509 solves the “secondary lighting” headache by seamlessly transferring lighting conditions across scenes — preserving realism without hallucination.

👉 Try it here

Upscale LoRA: Qwen-Image-Edit-2509-Upscale2K from starsfriday
No more pixelated outputs.This Upscale LoRA from starsfriday losslessly magnifies your generations to ~2K/4K resolution while preserving fine details — perfect for turning rough concepts into production-ready visuals.

🔗 Try it here

📬 Want More? Stay Updated.

Every week, we bring you:
● New model releases & upgrades
● AI research breakthroughs
● Open-source tools you can use today
● Community highlights that inspire

👉 Subscribe to The Tongyi Weekly and never miss a release.
Subscribe Now：https://www.linkedin.com/build-relation/newsletter-follow?entityUrn=7392460924453945345

Thank you for being part of this journey.

Tongyi Lab is a research institution under Alibaba Group dedicated to artificial intelligence and foundation models, focusing on the research, development, and innovative applications of AI models across diverse domains. Its research spans large language models (LLMs), multimodal understanding and generation, visual AIGC, speech technologies, and more.

Top comments (0)