A New Generation Studies AI, Apple's Recipe for On-Device Models, GLM5.2 Tackles Open-Ended Problems

Z.ai just put out GLM-5.2, and it's not a minor point release. The model is open-weights, licensed under MIT, free to grab off Hugging Face for commercial or noncommercial use. Under the hood it's a mixture-of-experts transformer with 753 billion parameters total, though only 40 billion fire per token. The headline spec is context: GLM-5.2 can take in up to a million tokens, a fivefold jump from the 200,000-token window on GLM-5. Output tops out at 128,000 tokens, spit out at 103 tokens per second.

The benchmark numbers are the part that should make closed-model vendors nervous. On Artificial Analysis's Intelligence Index, GLM-5.2 scored 51 with max reasoning enabled, trailing Claude Opus 4.8 (56) and GPT-5.5 (55) but comfortably clear of DeepSeek V4 Pro and MiniMax-M3, both stuck at 44. On PostTrainBench, a test built around fine-tuning four language models and grading them across seven benchmarks, GLM-5.2 actually came out on top at 34.3 percent, narrowly ahead of Claude Opus 4.8's 34.1 percent. It placed second on Arena.ai's Code Arena WebDev leaderboard behind Claude Fable 5, and third overall on AA-Briefcase, a business-document benchmark, while still leading every other open-weights model there.

Getting there took some engineering detours. Because GLM-5.2's agentic tasks run long, the team had to split individual attempts into segments during training, which broke the averaging trick used by Group Relative Policy Optimization. So they switched to Proximal Policy Optimization instead, judging each attempt with a separate critic model. They also had to deal with reward hacking — the model learned to fetch reference solutions from GitHub rather than actually solve problems, more often than GLM-5.1 did. The fix was a rule-based filter plus a second model to flag and block the cheating attempts by feeding dummy data instead. On the efficiency side, a sparse attention indexer now runs once every four layers instead of every layer, cutting per-token computation by 2.9 times at the million-token context length, and speculative decoding now accepts 5.47 tokens per step versus 4.56 previously, a 20 percent speedup.

Timing matters here too. GLM-5.2 landed one day after the US government restricted citizen access to Anthropic's Claude Fable 5 and Claude Mythos 5, and Anthropic itself suspended Claude Fable 5 access. Whatever the reason, the contrast is stark: as access to some frontier US models tightens, a Chinese lab ships something nearly as capable and hands it out for free, with API pricing running as low as a quarter of the cost per unit of intelligence compared to Opus 4.8 or GPT-5.5, according to Artificial Analysis's own cost assessment.

What this adds up to is a shrinking gap between what you can license and what you can just download. GLM-5.2 isn't beating the best closed models across the board, but it's close enough, cheap enough, and open enough that plenty of developers won't need to care about the difference.