z.ai released GLM-5.1, a 754-billion parameter model built for extended reasoning tasks. The model targets what the company calls "long-horizon" problems. Work that requires multi-step thinking rather than quick responses. Available on Hugging Face with Unsloth quantizations. Even the compressed IQ4_XS version sits at 361GB. You're not running this at home.
Early feedback paints a rough picture. People testing the model for coding ran into quantization artifacts, circular reasoning loops, and Chinese character glitches. One user called it "seriously gimped" for actual work. These aren't edge cases. They're fundamental reliability problems.
The release highlights a split in how teams tackle complex AI tasks. Cognizant's AI Lab has been developing MAKER, which spreads reasoning across millions of coordinated agents instead of stuffing everything into one giant model. Million-step reasoning with zero errors, they claim. GLM-5.1 bets the opposite, that raw parameter count can handle reasoning. The distributed approach looks more dependable right now.
What GLM-5.1 actually does well remains unclear. The long-horizon focus sounds promising. But if basic coding produces garbled output, those capabilities don't matter much. z.ai hasn't responded to user reports. That's a lot of compute delivering uncertain results.