GLM-4.7 and MiniMax M2.1 open up, Nvidia licenses Groq's tech

Christmas week is usually dead. This one had two serious open-weights releases and a deal that tells you where the chip money is heading.

GLM-4.7 and the open coding race

Zhipu released GLM-4.7 on 22 December, and the pitch is simple: better agentic coding, better tool use, weights you can download. I treat GLM releases as a barometer. Each version closes more of the gap to the closed frontier on the tasks I actually care about, which are long, multi-step jobs where the model has to keep its place.

MiniMax followed a day later with M2.1, the update to the M2 model that had already won a following among people running coding agents. It leans into multilingual code and longer agent loops.

Two things stand out when you put them next to each other. Both labs are shipping on a short cycle, a few months apart. And both are releasing weights, which means anyone can host them, fine-tune them, or run them inside their own network.

That last point is a big deal for regulated work. If a client will not let documents leave their environment, an open model on their own hardware is the only version of this technology they can say yes to.

Image editing that remembers

Qwen-Image-Edit-2511 improves consistency. Edit a photo of a person or a product several times and it now drifts less. Editing tools live or die on that, and it is a good sign that an open model is working on it.

Nvidia and Groq

Nvidia announced a non-exclusive licensing agreement with Groq for inference technology. I will not pretend to read the full deal from a press release. The takeaway I'm comfortable with is narrower: the company that dominates training chips is paying attention to fast inference, because that is where agents spend their time.

For a founder in the Gulf this matters in a concrete way. A model I can host myself means I can say yes to a client who insists that documents stay in the country, and I can say it without a long procurement of my own. The closed labs will keep the top of the leaderboard for now. The open ones are quickly becoming good enough for the large middle of real work, which is where most business value sits.

What I'd watch next: whether GLM-4.7 and M2.1 get picked up as the default open coding models in the tools developers already use.