New split in the AI world
The world of AI is changing fast. In mid-2026 we see two main ways.
One group of companies locks their models. The other group opens them to everyone. DeepSeek took a big step toward openness.
The company shared all code and weights of DeepSeek-V4. This means anyone can download and use it. The MIT license gives full freedom.
At the same time, Chinese company Zhipu.ai works on GLM-5.5. This model aims to match the best locked models.
AI w Biznesie helps firms pick the best tools. We show how to use open models at work.
DeepSeek DeepSpec – open code for everyone
DeepSeek just put out DeepSpec. This is a full set of tools to train and run models.
In the repository you find model weights, training code, and test scripts. Everything is under the MIT license.
This is important because other firms lock their models. For example, access to Fable 5 and Mythos 5 is stopped. GPT-5.6 also has limits.
DeepSeek does the opposite. It gives everyone access to its tech. This changes the rules of the game.
AI w Biznesie teaches clients how to use such open tools. We show how to set them up in a company.
What is DSpark?
DSpark is a new way to make models run faster. The method is called predictive decoding. It is a smart trick to speed up answers.
Normally a large model makes one word at a time. This is slow and costs a lot. DSpark works in a different way.
A small, fast model first writes a short draft text. Then the big model only checks if it is right. If the draft is good, the big model accepts it at once.
Old versions of this trick had a problem. After five or six words, the quality got worse. Each word did not look at the words before it.
DeepSeek added a light „head” to the fast model. This lets each word see what came before. They also added a scheduler. It decides how far the small model should guess. When the system is not busy, it guesses more. When it is busy, it guesses less.
Results in practice
DeepSeek tested DSpark on real user traffic. The results were impressive. For one user, speed went up by 60 to 85 percent.
The Flash version got the same boost. The Pro version also became 57 to 78 percent faster. The best part is that quality stayed the same. No loss at all.
The DeepSpec platform includes other methods too. You can compare DSpark with Eagle3 or DFlash. This is very helpful for researchers.
AI w Biznesie tests these tools for its clients. We show which solution works best for each task.
GLM-5.5 chases the best models
Good news comes from China. Model GLM-5.5 aims to match Claude 3 and Mythos. It is the next step after GLM-5.2.
The earlier version GLM-5.2 was already very strong. It beat GPT-5.4 and Gemini 3.1 Pro. It was equal to GPT-5.5 and Opus 4.8. In some tests it was even better.
Now Zhipu.ai works on a better model. GLM-5.5 wants to reach the level of Claude 3. This is a big goal.
The CEO of DeepSeek said that this is coming. In a talk with Elon Musk, he said it was only a matter of time. China has access to Huawei chips in large numbers. This lets them train models faster and cheaper.
Right now there is only a 14 percent chance that a Chinese model will be the best by the end of 2026. But the progress is very fast.
AI w Biznesie watches these changes. We help firms understand which models to use for their needs.
Why this matters
For a long time people thought open models would always be behind. Now this idea is changing. The gap is getting smaller.
GLM-5.2 already took the third place. It passed Gemini, which was widely used. Now OpenAI and Anthropic fight for first and second place. But Chinese models are getting closer every month.
By December 2026, GLM-5.5 could be in the top three. This is a big shift for the whole industry. It means users have more good choices.
AI w Biznesie gives advice on which model fits each job. You do not always need the most expensive one.
Why open source is winning
In 2026 more and more firms pick open models. There are several reasons for this trend. First, cost is much lower.
Models from locked labs are getting expensive. Open models you can run on your own machines. This saves money over time.
Second, reliability is key. Locked models can be limited at any moment. You might lose access without warning.
For example, Fable 5 was completely taken away. Some models ask for ID checks. Others block users from certain places. With an open model under MIT license, it works always. You run it on your own hardware.
AI w Biznesie helps firms set up open models. We show how to do it safely and cheaply.
Future belongs to openness
More developers choose open sources. They want control and long-term stability. Infrastructure costs are going down.
Running your own model is becoming money-saving. In the long run, this may be better than depending on locked firms. The risk is much smaller.
DeepSeek and Zhipu.ai show that openness works well. Their models are fast and low-cost. They prove that open does not mean low quality.
AI w Biznesie offers training for setting up open models. We teach you how to use them in your business. The world of AI is splitting into two sides. One locks doors, the other opens them. The choice is yours.
If you want to be independent, open sources are a good path. They give you full control over the tech you use. Follow our newsletter Universe of AI to stay updated on new changes.
Subscribe to the World of AI channel. Join the community on X under the name University of AIZ. See you in the next update. There will be even more news from the AI world.
No responses yet