Top AI Product

So Zhipu AI quietly released [GLM-5](https://huggingface.co/zai-org/GLM-5) on February 11th, and honestly, this thing deserves way more noise than it’s getting in the Western AI bubble. We’re talking about a 744-billion-parameter Mixture-of-Experts model with 40B active parameters per token, fully open-sourced under the MIT license. Yes, MIT — the “do whatever you want with it” license. And here’s the kicker: the entire model was trained on Huawei Ascend chips. Not a single NVIDIA GPU was involved.

That last point alone makes this a big deal. With US export controls squeezing China’s access to high-end NVIDIA hardware, Zhipu basically proved that you can train a frontier-class model on domestic silicon and still compete with the best. GLM-5 scored above 50 on the [Artificial Analysis Intelligence Index](https://artificialanalysis.ai/models/glm-5), making it the first open-weights model to ever cross that threshold. It sits comfortably alongside Claude Opus 4.5, GPT-5.2, and Gemini 3.0 Pro — not behind them.

The technical underpinnings are genuinely interesting too. Zhipu built a custom async reinforcement learning framework called [Slime](https://github.com/THUDM/slime) that decouples inference, evaluation, and parameter updates into parallel pipelines, killing the idle-time bottleneck that plagues standard RLHF training. They also integrated DeepSeek Sparse Attention to keep inference costs reasonable despite the model’s massive size. [VentureBeat highlighted](https://venturebeat.com/technology/z-ais-open-source-glm-5-achieves-record-low-hallucination-rate-and-leverages) its record-low hallucination rate, which tracks with the model feeling noticeably more grounded in my testing compared to other open alternatives.

The community response has been strong. [BuildFastWithAI did a solid writeup](https://www.buildfastwithai.com/blogs/glm-5-released-open-source-model-2026), the [GitHub repo](https://github.com/zai-org/GLM-5) already has deployment guides for vLLM, SGLang, and other inference stacks, and API access through Z.ai comes in at roughly $1.00 per million input tokens — significantly cheaper than comparable proprietary models. If you’ve been waiting for an open model that can genuinely hold its own against closed-source heavyweights, GLM-5 might be exactly what you’ve been looking for.