GLM-4.5: Coding Demos, Model Sizes and Z.ai’s Benchmarks

In July 2025, Matthew Berman examines GLM-4.5, an open-weight model from Z.ai, through coding demos and the company’s own benchmark charts. He argues that several Chinese open models are approaching leading proprietary systems, but the examples and reported scores are snapshots of particular tasks rather than proof of universal superiority.

The first demo is a 3D Rubik’s Cube simulator. Berman says he needed a couple of follow-up exchanges to fix a minor problem; the resulting interface rotates layers, scrambles the cube and displays move history. A standard cube appears to return to its solved state, followed by a 5×5 cube.

He also tries a 10×10 cube, but notices that its initial scramble is fairly limited. He asks the model to add a control for the number of scramble moves. The on-screen explanation describes solving by reversing the stored scramble sequence and inverting each move, so this shows an undo-based simulator rather than a solver independently finding a solution to an arbitrary cube state.

Next is a four-disc Tower of Hanoi. Berman asks the model to work through the recursive puzzle in its reasoning before creating a visualization, and says its displayed reasoning lists 15 moves. The generated visualization then animates a solution. This single small example does not resolve broader questions about models’ performance across puzzle sizes or conditions.

Berman asks for a ten-disc version at higher animation speed and moves on to an interactive Lego builder made with Three.js in a single HTML file. He can select colored bricks and stack them, but the first brick does not align perfectly with the base and another placement produces an error. He regards the result as promising rather than flawless.

His favorite demo is a 3D solar system with planet information and adjustable settings. The first attempt becomes stuck loading; after he tells the model about the problem, it revises the code and the visualization works. Berman performs this in the Z.ai chat interface’s canvas mode, not a local development environment.

Hovering over Venus reveals details such as its distance from the Sun and orbital period. Berman adjusts the simulation speed, orbits, labels, stars, planet size, distance scale and lighting. The demo shows interactive features, although the video does not independently verify every astronomical value.

Berman turns to the model release. GLM-4.5 has 355 billion total parameters, 32 billion active per token; the smaller GLM-4.5-Air has 106 billion total and 12 billion active. Both use a mixture-of-experts design and are released with weights under an MIT license, according to Z.ai’s model card.

Z.ai describes both as hybrid models with a thinking mode for complex reasoning and tool use and a non-thinking mode for immediate responses. In Berman’s quick tests, even asking California’s capital or requesting a simple story triggers visible thinking; he says he did not manage to activate the non-thinking behavior.

Berman shows Z.ai’s aggregate chart covering 12 benchmarks. The company reports 63.2 for GLM-4.5 and 59.8 for Air, compared with 65.0 for OpenAI o3 and 63.6 for Grok 4 in that chart. These are vendor-reported comparisons under its selected evaluation conditions, not a current or universal ranking.

The chart also groups agentic tasks, reasoning and coding. Berman notes GLM-4.5’s strong agentic results, a lower position in reasoning than several competing models, and a coding result behind the two Claude 4 models shown. Individual benchmark choices, test settings and scaffolding matter when interpreting these comparisons.

A second vendor chart plots SWE-bench Verified performance against total parameter count. Berman highlights GLM-4.5 near Kimi K2’s plotted score at a smaller total size, while Claude 4 Sonnet appears higher with its parameter count unknown. He notes that Z.ai used reinforcement-learning post-training for agentic capabilities.

Returning to the cube, Berman uses the newly added control to request 50 scramble moves on a 10×10 cube and then runs the solve animation. He does not count the moves independently. The visible description of the implementation indicates that it reverses recorded scramble moves, limiting what this result demonstrates.

He then tours examples provided by the model maker: a Flappy Bird-style game controlled with the space bar, a 3D maze, a movable to-do board, an SVG history of large language models, nested spinning hexagons and a Pokédex interface. These are selected demonstrations, not systematic tests of reliability.

Finally, the ten-disc Hanoi example presents a recursive algorithm and groups of steps rather than listing every move in the displayed reasoning. The animation finishes with 1,023 moves shown; the interface also displays 81 seconds and a speed setting of 10 moves per second. Berman praises the result, though an animation generated by code does not establish that the model manually reasoned through all 1,023 moves.

Berman concludes that open models have effectively caught up with closed models in the examples and benchmarks he has reviewed, while speculating that the then-unreleased GPT-5 might change the comparison. His judgment reflects the landscape and evidence available in July 2025.