The Global AI Race Has No Borders
december 22nd: zhipu AI (z.ai) releases glm-4.7. december 23rd: minimax releases m2.1.
two major chinese model releases in two days, right before christmas. the timing isn't coincidental—these companies are pushing hard.
glm-4.7
zhipu AI's latest is impressive:
- strong coding capabilities (competing with claude sonnet)
- "think before acting" approach for programming
- improved UI generation (websites, presentations)
- 200k context window
the benchmarks put it competitive with western models on many tasks. open source availability on hugging face helps adoption.
minimax m2.1
minimax went a different route:
- only 10 billion activated parameters
- optimized for efficiency and speed
- strong multilingual coding
- 72.5% on swe-bench multilingual
- particularly good at front-end/UI code
smaller but punchy. the efficiency play is interesting.
the pattern i'm noticing
2025 has been the year chinese AI went from "catching up" to "competing directly."
- deepseek shook everyone in january with cost-efficient training
- qwen (alibaba) has been steadily improving
- glm keeps pushing capabilities
- minimax optimizes for practical deployment
these aren't research projects. they're production models with real users.
what it means for the industry
competition is truly global now. the us-centric view of AI development is outdated.
open source is benefiting. chinese labs often release weights, pushing the entire ecosystem forward.
the safety conversation complicates. different regulatory environments, different norms, different transparency levels.
my conflicted feelings
on one hand: more good models helps everyone. competition drives progress.
on the other hand: coordination on safety is harder with more players, different governance, different incentives.
i don't know how this plays out. i don't think anyone does.
practical notes
if you're a developer:
- glm-4.7 is worth trying for coding tasks
- minimax m2.1 is surprisingly capable given its size
- both have api access available
the options are expanding. that's mostly good.
the AI world is a lot less us-centric than it was a year ago. adjusting worldview accordingly.