GLM 5.2

GLM 5.2 has been released, but is currently restricted to coding plan subscribers.

API and chatbot services for GLM 5.2 will launch next week.

GLM 5.2 will be open sourced under the MIT license.

The announcement post for GLM 5.2 is extremely short and lacks detailed information.

GLM 5.2 likely has the same parameter count as GLM 5.1 because it is an iterative 0.1 update.

GLM 5.1 was comparable to state-of-the-art models of its time.

GLM 5.1 is a 754 billion parameter mixture-of-experts model with 1 million context support.

If GLM 5.2 can perform well over long context lengths, it would be very valuable for an open model.

You can access GLM 5.2 by subscribing to the coding plan.

The coding plan subscription had lackluster performance in the past and recently increased its price.

The fact that GLM 5.2 will be open source makes the subscription price more reasonable.

The model's chain of thought shows detailed code checking, such as counting open and closed tags.

Watching the model's chain of thought gave confidence in the quality of the generated result.

The browser OS test generation took around 15.5 minutes in build mode.

The city game result had one of the most properly city-scaled maps the author has seen.

The generated game correctly implemented police pursuit mechanics (losing health on collision, decreasing wanted level when evading).

The game featured proper fog implementation with buildings appearing as the player approaches.

The Space Defender game did not work.

The Ether browser had a non-functional go button and a visually poor gradient, but was otherwise cool.

The terminal app was well done.

The hyperdrive 3D window feature was well implemented and smooth, rendering live windows in a WebGL scene.

The C++ skate game failed to render color on the boardwalk, which was a significant drawback.

Overall, the skate game was acceptable and pleasing for an open-weight model.

The C++ rally game using RayLib was exceptionally well done and impressive compared to other models.

The static subway scene was nicely done, with well-rendered tiles.

The model added extra adjustable settings (dust, flicker) beyond what the prompt required, going above and beyond.

The FPS game adaptation from the subway scene was well done overall, with an initial damage bug that was successfully fixed.

The initial watch model had incorrectly placed hands, falling short of expectations.

After receiving feedback, the watch model significantly improved, with correct hand positions and moving seconds.

The 3D printer simulation failed to extrude visible plastic and did not work.

The V8 engine STL model was very well done, print-ready, and impressive for an open-weight model.

The motorcycle racing game struggled with bike orientation and track layout, though some components were present.

The drum kit simulation had non-functional cymbals and hi-hat but was otherwise not bad.

Overall, GLM 5.2 is likely the best open-weight coding model currently available based on this testing.

The model's one-shot generations were not extremely impressive, but it improved significantly when given a chance to self-correct.

The model spends considerable time refining code after generation, similar to a meticulous barber.