GLM 5.2

GLM 5.2 is the new number one open-source model and the gap to other open models is insane.

GLM 5.2 beats the best GPT and the best Gemini across multiple benchmarks.

GLM 5.2 beats GPT 5.5 and Gemini 3.1 Pro by a huge margin on long autonomous coding benchmarks and is close to Claude Opus 4.8.

To unleash GLM 5.2's full potential, users should use agentic frameworks like OpenClaw, Hermes, Claude Code, or ZAI's Zcode.

GLM 5.2 required additional prompting and about 15 minutes to build a complex 3D interactive Earth, while Claude Fable 5 completed it in one prompt, but the result is still extremely impressive for an open-source model.

GLM 5.2 outperformed Claude Fable 5 on a mechanical watch 3D animation task, succeeding where Claude Fable 5 could not get it correct.

GLM 5.2 built a ray tracing simulation from scratch without external libraries, performing better than Claude Fable 5 in terms of latency and usability.

GLM 5.2 created a promo video with TTS voice-over and animation using an external framework without any additional prompting, and generally requires very minimal handholding.

GLM 5.2 generated a 3D animated V8 engine in one prompt in about 13 minutes.

GLM 5.2 can compose a 32-bar song but the result lacks complexity, won't win a Grammy, and is similar to what Claude Fable 5 generated.

GLM 5.2 lacks native vision capabilities and failed to correctly find a hidden frog in an image, relying on an external vision model which gave a wrong answer.

GLM 5.2 excels at deep research, generating concise, well-structured reports with tables, flowcharts, and detailed visualizations.

GLM 5.2 features a 1 million token context window.

GLM 5.2 is a 753 billion parameter model, 1.51 terabytes in size, and is not feasible to run on consumer devices.

GLM 5.2 is released under the MIT open-source license, making it very permissive.

GLM 5.2 scored 46.2 on the Deep Suite software engineering benchmark, the highest among any open model.

GLM 5.2 outperforms the best GPT and Gemini on the Humanity's Last Exam benchmark, indicating greater obscure scientific knowledge.

GLM 5.2 is by far the best open model on the RuneScape benchmark.

GLM 5.2 uses an index share architecture that reduces compute by 2.9 times.

GLM 5.2 incorporates an improved MTP layer that increases decoding length by up to 20%, boosting generation efficiency.

GLM 5.2 beats Claude Opus 4.8 in front-end coding on the Frontier leaderboard and even outperforms Claude Fable 5 in design on Design Arena.

Open-source models like GLM 5.2 provide sovereignty and privacy because they can be self-hosted, keeping data under user control unlike closed-source APIs.