ZAI has released GLM 5.2, now the best open-source AI model, outperforming GPT and Gemini on multiple benchmarks. It features a 1 million token context window and an MIT open-source license.
Availability and Agentic Frameworks ⏱ 0:32
•Available for free at chat.z.ai with GLM 5.2 selection.•Best used with agentic frameworks: OpenClaw, Hermes, Claude Code, or ZAI's own Zcode (free, supports Mac, Windows, Linux).•Zcode is similar to Codeex, allows multiple projects and files.Performance Demos ⏱ 1:33
•3D Digital Twin of Earth: Built interactive 3D Earth with zoom, city stats, toggles for cloud cover, flight traffic, day/night, night city lights. Required ~15 minutes and additional prompting; Claude Fable 5 did it in one prompt.•Promo Video Creation: Created a promo video for Wise using Gemini TTS and hyperframes framework. Took ~20 minutes, used ~100k tokens, cash hit rate provided but omitted in transcript.•3D V8 Engine Model: Generated in 13 minutes with one prompt, single HTML file, working explosion and speed controls. Used ~28,000 tokens, 96% cash hit rate.•Mechanical Watch Model: More challenging; required error fixing and enhancements. Gears had alignment issues, but Claude Fable 5 also failed. Used ~80,000 tokens, 94% cash hit rate.•Ray Tracing Simulation: Built from scratch without 3JS, with adjustable parameters for sphere, cube, pyramid. Took ~20 minutes; worked better than Claude Fable 5.•Music Composition: Created a 32-bar song with piano roll interface. Required minor fixes; lacked complexity and auto-panning. Similar quality to Claude Fable 5.•Manim Butterfly Animation: Used Manim package; required installation. Produced butterfly with 12 circles; after refinement, added more details. Total time ~22 minutes for final version.Online Chat Interface Tests ⏱ 22:00
•Vision Test: Uploaded frog image; GLM 5.2 lacks native vision, used external VLM. Incorrect answer after ~10 minutes.•Deep Research: Prompt on leukemia molecular drivers; generated detailed response with tables, flowchart, timeline, drug table, web chart. Concise and thorough.Specifications ⏱ 24:03
| Feature | Detail |
|---|
| Context window | 1 million tokens (~700,000 words or small-to-medium codebase) |
| License | MIT open-source |
| Parameters | 753 billion |
| Size | 1.51 terabytes |
Benchmarks and Leaderboards ⏱ 24:43
| Benchmark | Performance |
|---|
| Long autonomous coding tasks (Frontier Suite, Post Train Bench) | Beats GPT 5.5 and Gemini 3.1 Pro by huge margin; close to Opus 4.8 |
| SWEBench Pro | Beats best GPT and Gemini; close to best Claude |
| Terminal Bench | Beats best GPT and Gemini; close to best Claude |
| Deep Suite | Score 46.2, highest open model on leaderboard |
| Humanity's Last Exam | Beats best GPT and Gemini |
| Frontier leaderboard | Close to Opus 4.8, beats GPT 5.5 |
| Front-end coding | Beats Opus 4.8 |
| Design Arena | Beats Claude Fable 5 and other closed models |
| Runescape Bench | Best open model by far |
Architecture and Open-Source Value ⏱ 26:05
•Architecture tweaks: index share (reduces compute by 2.9x), improved MTP layer (increases decoding length by up to 20%).•Fully open weights under MIT license on Hugging Face.•Open-source advantages: not dependent on closed labs (e.g., Anthropic gatekeeping, US government ban on latest Claude), privacy (data stays local), community fine-tuning and improvements.Key Takeaways
•GLM 5.2 is the top open-source model, beating GPT 5.5 and Gemini 3.1 Pro on multiple benchmarks.•It supports a 1 million token context window and is MIT-licensed.•The model excels in coding, front-end design, and deep research tasks.•It can be used via free chat interface or agentic frameworks like Zcode, Claude Code, OpenClaw, or Hermes.•Architecture improvements include index share (2.9x compute reduction) and improved MTP layer (20% longer decoding).•Open-source nature ensures user sovereignty, privacy, and community-driven enhancements.Conclusion
GLM 5.2 is a groundbreaking open-source AI model that rivals and often surpasses closed-source competitors. ZAI continues to lead in open-source AI development.
