A new version of the GLM4.7-Flash language model, called REAP, is available, promising superior performance in the field of agentic coding.

GLM4.7-Flash REAP: focus on coding

This version, obtained through REAP pruning techniques at 25%, has been specifically calibrated to improve agentic coding capabilities. Initial results, obtained through the SWE-rebench benchmark, show a significant step forward compared to the GLM4.6 model, with remarkable efficiency in relation to model size.

Feedback and future developments

The development team is evaluating the possibility of releasing REAP versions dedicated to creative writing, in response to user feedback that has found a negative impact on multilingual and creative writing skills with the current versions. Evaluations are underway to identify the most suitable datasets for this purpose. MiniMax-M2.1 is also under development, which should be positioned between the 25% and 40% REAP versions.