# Qwen3-Omni and VL arrive, Gemini Robotics 1.5 learns to think first

> Week of 22 Sept 2025: Qwen3-Omni, Qwen3-VL, DeepSeek V3.1 Terminus, HunyuanImage 3.0, Gemini Robotics 1.5, Meta CWM, GDPval.

Published 13 Oct 2026. Coverage: 22 Sept 2025 to 28 Sept 2025. By Ali Asif.

## Items

1. **Qwen3-Omni 30B-A3B Instruct** (Alibaba Qwen, Open weights): One open model takes text, images, audio and video in and talks back with speech. [Primary source](https://huggingface.co/Qwen/Qwen3-Omni-30B-A3B-Instruct)
2. **Qwen3-VL 235B-A22B Instruct** (Alibaba Qwen, Open weights): The largest open Qwen vision-language model so far, aimed at documents, video and GUI agents. [Primary source](https://huggingface.co/Qwen/Qwen3-VL-235B-A22B-Instruct)
3. **DeepSeek-V3.1-Terminus** (DeepSeek, Open weights): A refinement of V3.1 that fixes mixed-language output and improves its code and search agents. [Primary source](https://huggingface.co/deepseek-ai/DeepSeek-V3.1-Terminus)
4. **HunyuanImage 3.0** (Tencent, Image, video, world): A very large open image model that generates in the same autoregressive framework as a language model. [Primary source](https://huggingface.co/tencent/HunyuanImage-3.0)
5. **Qwen-Image-Edit-2509** (Alibaba Qwen, Image, video, world): Better multi-image editing and more consistent faces and text in edited pictures. [Primary source](https://huggingface.co/Qwen/Qwen-Image-Edit-2509)
6. **Gemini Robotics 1.5** (Google DeepMind, Robotics): A robot model that plans in natural language before acting, paired with an embodied reasoning model that can call tools. [Primary source](https://deepmind.google/discover/blog/gemini-robotics-15-brings-ai-agents-into-the-physical-world/)
7. **Code World Model (CWM)** (Meta FAIR, Science): A 32B research model trained on how code executes, not only on how it reads. [Primary source](https://huggingface.co/facebook/cwm)
8. **GDPval** (OpenAI, Policy and business): A benchmark built from real work products by experts across 44 occupations, to measure models on economically useful tasks. [Primary source](https://cdn.openai.com/pdf/d5eb7428-c4e9-4a33-bd86-86dd4bcf12ce/GDPval.pdf)

Alibaba had a very loud week, and I'll start there because the scale of it is hard to miss.

## The Qwen flood

Qwen3-Omni is one open model that accepts text, images, audio and video and answers in text or speech. Qwen3-VL, in a 235B mixture-of-experts size, is the vision-language model I'd test first for document work. Think scanned supplier certificates, drawings with stamps, tables that don't survive OCR. Qwen-Image-Edit-2509 tightened up multi-image editing. That's three serious open releases in a few days, and the cadence itself is the story. Qwen3-Omni in particular matters for voice work: one open model that listens and speaks removes a pipeline of three separate services. If you plan a product around one lab's roadmap, you are betting against how fast this one moves.

DeepSeek put out V3.1-Terminus, a cleanup of V3.1 that fixes stray language mixing and improves its agent behavior. Tencent opened HunyuanImage 3.0, a big image model built on an autoregressive design instead of the usual diffusion approach.

## Robots start to think out loud

Google DeepMind launched Gemini Robotics 1.5 and Gemini Robotics-ER 1.5. The model writes out a plan in language before it moves, and the ER model reasons about the physical scene and can call tools like search. I work on aviation lighting and airfield infrastructure, and I'm not near a humanoid in my day job. Still, a robot that explains its next step is a robot a safety officer can supervise, and that is the bar in my industry.

## Papers and measuring work

Meta released CWM, a 32B research model trained on traces of code execution, which is a different bet than simply reading more code. If it works, it could produce models that understand what a program does and not only what it looks like. OpenAI released GDPval, a benchmark of tasks built from real professional work across 44 occupations, graded by experts. It is the closest thing so far to a test of whether a model can do a job's output rather than answer a quiz.

## What I'd watch next

Everyone seems to be waiting on an Anthropic release, and I'd be surprised if the month ends without one. Claude Sonnet 4 is four months old and the coding competition has moved. I'd be surprised if the month ends without one.

## My take

GDPval is the benchmark I'd rather see more of. It scores models on deliverables that professionals produce, like a memo or a schedule, and not on puzzle questions. Our work at a company like mine is exactly those deliverables, so I'll be running my own small version of it on bid and procurement tasks.

## Sources

- https://huggingface.co/Qwen/Qwen3-Omni-30B-A3B-Instruct
- https://huggingface.co/Qwen/Qwen3-VL-235B-A22B-Instruct
- https://huggingface.co/deepseek-ai/DeepSeek-V3.1-Terminus
- https://huggingface.co/tencent/HunyuanImage-3.0
- https://huggingface.co/Qwen/Qwen-Image-Edit-2509
- https://deepmind.google/discover/blog/gemini-robotics-15-brings-ai-agents-into-the-physical-world/
- https://huggingface.co/facebook/cwm
- https://cdn.openai.com/pdf/d5eb7428-c4e9-4a33-bd86-86dd4bcf12ce/GDPval.pdf

Canonical: https://heyaliasif.com/ai/2025-09-22-week-in-ai