Breaking

Quantized GLM-5.2 builds solvable Rubik's Cube web app in one shot

July 2, 2026 at 11:35 EDT

Foundation Models
Open Source
Software Dev & Coding

A community-quantized, pruned build of Z.ai's open-weight model GLM-5.2, dubbed "GLM-5.2-REAP-NVFP4," has drawn attention among developers after a demo showed it generating a fully working HTML app featuring a 3D Rubik's Cube and an auto-solve button from a single prompt.

Late June 2026 · GLM-5.2-REAP-NVFP4

A pruned, quantized open model builds a working 3D Rubik's Cube in one shot

From a single prompt — "build a rubik's cube with a button that solves it, in HTML" — a compressed version of GLM-5.2 generated interactive HTML with real scrambles, live state updates, and an auto-Solve button, no follow-up instructions needed.

744B

Total params in the original MoE model

469.2B

After REAP pruning + NVFP4 — 61.4% retained

Token context window supported

Compression: shrinking the giant to run locally

Total parameters — original vs. pruned & quantized (same scale)

744B

Original GLM-5.2

469.2B

REAP + NVFP4

Two techniques combined: REAP prunes low-activation MoE experts · NVFP4 applies 4-bit quantization

What holds up ✓

Retains ~69–81% of the original's benchmark level
One-shot HTML / Three.js prototyping stays strong
Runs on 4×RTX PRO 6000 Blackwell (96GB each)

The trade-offs ⚠

REAP pruning degrades some knowledge recall
Original remains superior on large factual tasks
Still needs multiple high-end GPUs — a hurdle for most

The bigger picture

Combining pruning and quantization to bring 700B+ open MoE models onto on-hand hardware is spreading to the DeepSeek family and Qwen MoE. This one-shot demo — a small, local model matching closed large-scale visual generation — is seen as a symbol of that practicality.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.