DeepSeek's mHC paper, Qwen-Image-2512 and a Tencent translation model

The year ended quietly on the surface. Underneath, the open-model pipeline did not take a day off.

DeepSeek's mHC

The item I would hand to a smart friend first is a paper. On the last day of the year DeepSeek posted "mHC: Manifold-Constrained Hyper-Connections". The short version: transformers pass information forward through residual connections, and earlier work called hyper-connections tried to make those paths wider and richer. That helped capacity but made training unstable at scale. DeepSeek's fix constrains the connections so the signal stays well behaved.

I am an operator, not a researcher, so I am reading this as a signal and not as proof. What the paper says to me is that a lab with limited access to the newest chips keeps finding efficiency gains in the architecture itself. That has been DeepSeek's pattern for a while, and it is why I pay attention to their output even when the headline is not a new chatbot.

Image and translation

Alibaba shipped Qwen-Image-2512, an update to its open text-to-image model. The notes focus on more natural-looking people and cleaner text inside images, two things that image models have struggled with for years. Having that in open weights means anyone can run it privately.

Tencent released HY-MT1.5, with a 1.8B version small enough for a phone. Translation is the kind of task where a small, fast, local model is often the right answer. For anyone working across Arabic and English documents every day, I want more of these and fewer calls to a big hosted model.

A coding model to test, not to trust

IQuest-Coder-V1 appeared with eye-catching coding claims. I have a rule for any new open coding model: I run it on my own tasks before I believe a leaderboard number. Treat this one the same way.

Years end with a lot of recap posts, and I skimmed several. Most of them list the same dozen launches. What I took from the year was less about any single model and more about the cadence: a visible jump every few months, and a falling price for the previous generation. If you plan a business around today's model quality you will be wrong in a quarter, so plan around the trend, and build so that swapping the model is a configuration change.

What I'd watch next: whether the mHC idea shows up in a DeepSeek model, which would confirm it works at real scale.