We let the hot Chinese models play chess
Databricks has been a partner of all the frontier models since the beginning. We are used to running codex by OpenAI and Claude by Anthropic. However, recently there have been some really hot Chinese models catching up with the US models. They are Kimi by moonshoot AI and GLM by Z.ai.
It’s exciting that Databricks is a Day 0 launch partner for these models, which means we don’t need to maintain 10 different subscriptions and can leverage them out of the box from Databricks. The advantage for enterprises really means that their security team doesn’t need to worry about securing different endpoints and the risk of having employees download internal data and upload it to different websites.
- Data stays governed by Unity Catalog, and the RLS and CLS are intact
- Use Unity AI Gateway to control costs and access to models
- Use Omnigent for vibe-coding tasks right within Databricks
As we all know, the agents are very capable of coding. However, as the AGI discussion heats up, I want to find out how these new reasoning agents reason on some ancient problems, which is chess. The goal here really isn’t to beat a chess engine, but to determine if they can approach a “new” problem and reason it out without instructions. So I chose to let them play chess again. For simplicity, I only played Kimi K3 against GLM 5.3 Flash.
Throughout the development, there were a few observations:
- LLMs still don’t know how to play chess
The game was originally designed using Claude code. However, GLM constantly exceeded the token limits. A little digging revealed that the prompt was instructing the LLM to reason every move again. - GLM takes longer to reason
Despite giving the model 5 minutes per model and instructing it to be quick, GLM would still take its time to think. So it can never go beyond 10 moves despite the 5-minute allowance. - It memorizes openings well
Perhaps this is the biggest strength of LLMs. It reads all books and knows all the textbook knowledge. But without going into the middle game, it is hard to tell which model is a stronger chess player.
Conclusion and next steps
The Chinese models have arrived Databricks. And it’s exciting that we can also have a choice of these open-weight models. They are proven to be cost-effective and have a high completion rate, according to a recent blog post by Databricks.
Get Jason Yip’s stories in your inbox
Join Medium for free to get updates from this writer.
As we can see from the diagram below, GLM 5.2 is on par with Claude Opus 4.8, but cheaper.
There are tremendous opportunities ahead with the Databricks AI stack now coming with Omnigent and Chinese coding models. And I am sure Databricks will continue to support all frontier models regardless of where they come from. And enterprises can safeguard them using Unity AI Gateway.
The next step in this chess game is perhaps to have the models play against a weakened stockfish chess engine to see how many moves they will survive. Stay tuned for more chess games!
