Z.AI Model Releases, Tracked
17 releases across 6 models tracked from June 16, 2026 to September 2, 2026. Updated as new Z.AI releases are covered.
- News BriefGLM-5.3-FlashZ.AI Launches GLM-5.3-Flash: 320B Multimodal Open-Weight MoE Unveils Identity Behind Ox Alpha
Z.AI has officially launched GLM-5.3-Flash under the MIT license with 320B total parameters and 18B active parameters, confirming it was the mystery model previously tested as Ox Alpha.
- Hands-on TestOx AlphaI Prompted Ox Alpha to Build a 3D Open-World Survival Game — The Code Worked First Try, the Menu Took a Second Click
The unannounced stealth model Ox Alpha generated a complete 3D open-world survival game in a single 738-line HTML file — procedural terrain across four biomes, a full crafting and placement system, and a day/night cycle, all on the first attempt. The only snag was a pointer-lock click that needed a second try. Full prompt and live demo inside.
- Hands-on TestOx AlphaI Prompted Ox Alpha to Build a Top-Down Zombie Wave Survival Game — Spatial-Grid Pathfinding for 40+ Enemies, One Shot
Ox Alpha built a complete top-down 3D zombie wave-defense shooter in a single HTML file — spatial-grid neighbor queries for enemy separation, object pooling for zombies and blood effects, three enemy types, and a between-wave upgrade card system. One-shot, no fixes needed.
- Hands-on TestOx AlphaI Prompted Ox Alpha to Build a 3D Voxel Island Explorer — One Bug Found, One Prompt to Fix It
Ox Alpha built a complete first-person voxel island game in a single HTML file — Simplex noise terrain, mine/place block interaction, a day/night cycle, and generative pentatonic ambient music. The one bug we hit — grass and dirt blocks rendering the same color in the world — was fixed with a single follow-up prompt.
- Hands-on TestOx AlphaI Prompted Ox Alpha to Build a 3D Marble Physics Maze — Custom Collision Math, No Physics Library, One Shot
Ox Alpha built a complete 3D marble-tilt maze game from scratch in a single HTML file — hand-written sphere-box collision physics, hole-suction pull, rotating hazard bars solved in local rotated space, and three full stages with checkpoints, boosters, and bounce pads. No physics engine, no fixes needed.
- Hands-on TestOx AlphaI Prompted Ox Alpha to Build a Cyberpunk Hoverbike Endless Runner — One Shot, Full Synthwave Soundtrack Included
Ox Alpha built a complete 3D cyberpunk hoverbike endless runner in a single HTML file on the first try — glassmorphism HUD, a procedurally generated synthwave bass loop, banking lane-shift physics, and a full pooled obstacle/pickup system. No fixes needed.
- Deep DiveOx AlphaOx Alpha Debuts on OpenRouter: Mystery 1M-Context Model Sparks Unverified Community Benchmark Claims
An unannounced model named Ox Alpha appeared on OpenRouter and OpenCode with a 1M token context window and free access. Unverified community-run tests claim strong coding benchmark results, though nothing has been independently confirmed.
- News BriefGLM-5.2 TurboZ.AI Releases GLM-5.2 Turbo: High-Speed Open-Weight MoE for Production Agent Loops
Z.AI has launched GLM-5.2 Turbo, an inference-optimized 744B open-weight MoE model with a 1M token context window under the MIT license, delivering lower latency for high-frequency coding agent workflows.
- News BriefOx AlphaOx Alpha Launches on OpenRouter: 1M-Context AI Model Available in Free Preview
An unannounced AI model named Ox Alpha has appeared on OpenRouter and OpenCode with a 1M token context window, multimodal capabilities, and free tier access.
- Deep DiveGLM-5.3Deep Dive: Z.ai's GLM-5.3 Pushes MoE Reasoning Boundaries with Scaled RL Post-Training
Z.ai released GLM-5.3 on August 14, 2026, delivering massive benchmark improvements in terminal automation, coding, and cybersecurity. We analyze the scaled reinforcement learning that drove these post-training gains on the 743B MoE architecture.
- News BriefGLM-5.3Z.ai Releases GLM-5.3 MoE Model with Scaled Post-Training and Cybersecurity Upgrades
Z.ai released GLM-5.3 on August 14, 2026 — a 743-billion parameter Mixture-of-Experts model. While sharing the same base architecture as GLM-5.2, scaled post-training pushes its Terminal-Bench 3.0 performance from 4.6% to 28.3%.
- Deep DiveGLM-5.2GLM-5.2 Can Do Nearly Everything a Frontier Model Can. SaferAI Says It Has Almost No Guardrails.
SaferAI's independent evaluation of Z.ai's GLM-5.2 found the model matches GPT-5.5 and Claude Opus 4.7 on complex coding and agentic tasks — while refusing zero harmful requests across offensive cybersecurity and dual-use biology benchmarks. Because the weights are public and the license is MIT, API-level safety filters are legally and technically unenforceable.
- Hands-on TestGLM-5.2I Prompted GLM-5.2 on Max Effort to Build a 3D Open-World Survival Game — It Nailed It in One Prompt
Z.ai's 753B open-weight GLM-5.2 generated a complete 3D open-world survival game in a single 1,014-line HTML file on the first try — procedural Perlin noise terrain, three biomes, a 120-second day/night cycle, resource harvesting, and placeable crafting structures. No follow-up prompts, no fixes. Full prompt and live demo inside.
- Deep DiveGLM 5.2The Alliance That the Hugging Face Hack Built: NVIDIA and 37 Companies Are Betting Open AI Wins the Security War
NVIDIA and 37 companies launched the Open Secure AI Alliance on July 27, arguing that open-weight AI is essential for cybersecurity defense. The founding trigger was an AI-driven breach of Hugging Face's production systems — and the open model that helped contain it was GLM 5.2, a Chinese open-weight that closed-source tools couldn't replicate.
- Hands-on TestGLM 5.2I prompted GLM 5.2 Deep Think Maxx to build a 3D zombie survival game — 2,400 lines of Three.js in one shot
GLM 5.2 Deep Think Maxx generated a complete 3D zombie survival fighting game in a single HTML file — third-person camera, wave system, combo combat, and procedural city. Full prompt and breakdown inside.
- Hands-on TestGLM 5.2I prompted GLM 5.2 Deep Think Max to build a Retro Pseudo-3D Raycast FPS Game — 1,521 lines of vanilla JS in one shot
We challenged GLM 5.2 Deep Think Max to build a complete Wolfenstein-style pseudo-3D FPS in a single HTML file — DDA raycasting, three weapons, wave-based AI, and fully synthesized WebAudio sound, with zero libraries and zero external assets. Full prompt, code breakdown, and an honest look at what shipped half-finished.
- News BriefGLM-5.2GLM-5.2 released: Z.ai's 753B open-weight model matches GPT-5.2, draws a federal safety assessment
Z.ai (formerly Zhipu AI) released GLM-5.2 — a 753B parameter open-weight MoE coding model with a 1M-token context window under the MIT license. A NIST assessment later found its capabilities rival GPT-5.2, with mixed safety results.