Tencent's new Chronicles-OCR benchmark reveals that top AI models like GPT-5 and Gemini fail miserably at recognizing ancient Chinese scripts, scoring near zero on oracle bone and bronze inscriptions.
Zhipu AI's founder predicts AI models will master long-horizon tasks and evolve to self-train, with Claude potentially already doing so — and a massive 2 million chip cluster on the horizon.
Zhipu AI releases GLM-5V-Turbo, its first multimodal programming model, beating Claude Opus on Design2Code benchmark with a score of 94.8. The model integrates vision from pre-training, not as an afterthought.
Zhipu AI explains the root cause of garbled output in its GLM-5 coding agent model: two race conditions in the KV cache, plus a trick for anomaly detection using speculative decoding acceptance rates.
ByteDance, Zhipu AI, and Alibaba land on TIME's first-ever AI-focused list, alongside OpenAI, Anthropic, and a surprise 'no-Western-chips' label for Zhipu.
The Chinese AI company is giving developers more time to take advantage of its GLM-5.1 and GLM-5-Turbo models at triple the usual usage limit through June 30.
Zhipu AI is ending auto-renewals for its unlimited weekly coding plan on April 30, citing surging demand. Affected users get two months of the new plan free.
Cheap AI coding plans are bait: companies want your private code data, not your subscription fee. Violate the 'code-only' rules, and you'll get banned without a refund.
